If you’re not a subscriber, here’s what you missed this week:

1. Claude Opus 4.8: The Complete Prompting Playbook for 2026 🧠 [How to get the best results from Anthropic’s most powerful model every single time, including the new Dynamic Workflows feature]

2. Claude for Small Business: Setup, Workflows, and the Prompts That Make It Work 💼 [The complete guide to Anthropic’s Claude for Small Business - covering setup, all 30 capabilities, 8 integration deep-dives, and the exact prompts that turn Claude into your most productive employee]

3. SpaceX IPO: Why the Biggest Listing in History Might Be Uninvestable at $1.75 Trillion 🪐 [The S-1 reveals a $7 billion Starlink profit machine subsidizing $30B in annual AI cash burn, while Elon Musk holds 85% of the vote with no sunset, no independent board, and no path to remove him]

4. The Claude Finance Playbook: How CFOs Use AI to Build Models, Forecast Cash, and Close Books Faster 📊 [A practical guide to setting up Claude for real finance work, with battle-tested prompts optimized for Claude Opus 4.7]

5. Ex-OpenAI’s Aschenbrenner Bet $5.5B on AI Chips. Now He’s Betting $8.5B Against Them 💸 [Leopold Aschenbrenner’s AI hedge fund Situational Awareness went from pure AI long to 62% short semiconductors in one quarter. Here’s what changed and why the thesis still holds]

6. Monzo’s 2026 financials: the £1.7 billion revenue machine that still has to prove it can escape the UK’s Orbit 🏦🇬🇧 [breaking down the unit economics that matter, the Revolut comparison investors keep getting wrong, the governance drama, IPO chances & what’s next for Monzo + bonus deep dive into Revolut’s latest financials & their foundational AI model for money inside]

7. Starling Bank’s 2026 financials: 2.5x Monzo’s profit on half the revenue, and a £70M SaaS pipeline that changes the IPO math 📈🧠 [unpacking the most important financial facts & figures, how it stacks against Monzo & Revolut, & why Starling’s balance sheet is a fortress + bonus deep dives into the latest financials of Monzo & Revolut inside]

8. Coatue’s May 2026 Report: The $12 Trillion AI Bet and Who Ends Up on the Wrong Side 📊 [Coatue’s latest public markets report reveals the most extreme winner/loser split in tech stock history, a $12 trillion AI capex wave, and one brutal framework that predicts exactly who wins]

9. Anthropic Just Told AI Founders Exactly What to Build in 2026 🦄 [1 million conversations. 9 consumer AI domains. A full founder playbook - plus where Anthropic’s own products will and won’t compete]

10. The Complete Claude /goal Guide for AI Agents 🤖 [What Claude /goal actually is, why most invocations fail, and how to write production-grade goal conditions for financial research, backtesting, and live market monitoring]

11. How to Build an AI Agent from Scratch (With Working Code) 🤖 [The design framework, working code, and hard-won lessons that take you from “I want to build an AI Agent” to a working one - in under 60 minutes]

12. Claude Cowork Commands, Scheduled Tasks & Automation Workflows: The Operator’s Playbook 🤖 [Every real slash command, the exact prompts that run deal flow and investor updates on autopilot, and the failure modes that will cost you hours if you don’t know them first]

As for today, here are the 2 fascinating FinTech stories that are changing the world of financial technology as we know it. This was yet another wild week in the financial technology space, so make sure to check all the above stories.

Your next broker is a chatbot 🤑🤖

The news 🗞️ The gap between “AI can analyze markets” and “AI can execute trades” just closed, and it happened inside a chatbot.

Let’s take a look.

More on this 👉 Liquid Co-Invest, built by ex-Two Sigma quant Franklyn Wang, ships as a ChatGPT App Store plugin and Claude connector that turns a conversation into a full trading workflow.

Fund an account, pull institutional-grade analysis (on-chain flows, funding rates, liquidation maps, macro context with sourcing), build a multi-asset portfolio, and execute live trades.

→ Crypto perps with leverage, US equities, FX, commodities, Polymarket positions, and pre-IPO secondaries in names like OpenAI, Anthropic, and SpaceX.

→ Over 500 markets, available in all 50 states, setup in about a minute.

→ Non-custodial, so funds stay in the user’s wallet, and the AI can’t move money without an explicit tap on a confirmation card showing exact position details.

Zoom out 🔎 Looking at the bigger picture, the smart observation here isn’t about the interface. It’s about where the moat sits.

Liquid doesn’t own the AI. OpenAI and Anthropic do, and every model upgrade makes Liquid’s product better at zero marginal cost. What Liquid owns is the hard, unglamorous infrastructure underneath: deep venue integrations across Hyperliquid, Lighter, and Ostium, regulatory positioning that sidesteps custody headaches, and access to private-market secondaries that most retail investors can’t touch. And that’s exactly what Anthropic itself recently told AI founders to build:

Most importantly, they’re not competing with the likes of Robinhood, Coinbase, etc. They’re making their interface feel like a fax machine 😬

The incumbents won’t sit still, of course.

Robinhood already has the regulatory licenses, 24 million funded accounts, and a mobile-native DNA that rewired retail trading once before. Coinbase sits on the deepest institutional crypto rails in the US and has been quietly building agent-friendly APIs.

Both will almost certainly ship their own LLM connectors or acquire someone who has. Their real advantage is custody: they already hold the assets. But that’s also their constraint. Building conversational execution on top of legacy brokerage architecture is a different engineering problem than building it natively, and incumbents tend to bolt on what insurgents build in.

The race therefore now is whether Robinhood and Coinbase can move fast enough to own the agentic layer themselves, or whether they get relegated to dumb pipes that infrastructure players like Liquid route through.

THE TAKEAWAY ✈️

What’s next? 🤔 First and foremost, we must note that the risk profile is real here, though. Leverage in the hands of someone who has just learned to prompt is a rough combination. Herding effects from thousands of users feeding similar queries into the same models could amplify volatility rather than dampen it. And regulators haven’t even begun to wrestle with liability when an AI’s suggestion leads to a blowup… Looking ahead, I’d watch two things. First, whether traditional brokers scramble to build their own LLM connectors or try to acquire the infrastructure layer. Second, the pre-IPO access angle. If conversational AI makes illiquid private secondaries feel as tradable as spot crypto, that reprices how late-stage companies think about liquidity timelines entirely. The first-order story is a better trading UI. The second-order story is a new distribution channel for capital markets, and that’s the one worth tracking. Either way, this is a perfect glimpse into the future of agentic finance.

ICYMI:

Catena Labs wants a bank charter for AI Agents, and the OCC said “We’ll Listen” 🏦🤖

The news 🗞️ The first serious regulatory test for agentic finance is now on the clock.

Catena Labs, founded by Circle co-founder Patrick Neville, emerged from stealth last week with a $30M Series A from Acrew Capital and a16z crypto, bringing total funding to $48M.

But the money isn’t the story here.

The story here is that the OCC accepted Catena’s filing for a national trust bank charter - a de novo institution designed from scratch to let AI agents custody assets, settle payments, and manage money under federal supervision.

Let’s take a quick look to see what it’s all about, why it matters, and most importantly - what’s next.

More on this 👉 Every major player in payments is now racing to build agent infrastructure.

→ Stripe launched its Agentic Commerce Protocol with OpenAI, then joined Google’s UCP.

→ Coinbase shipped AgentKit and x402.

→ Google partnered with Coinbase and AmEx on agent payments.

All of them are API layers and protocols that ride on existing banking relationships. Catena is the only one trying to become the bank. That’s a slower path, but if the charter clears, it’s a moat that no SDK can replicate.

Zoom out 🔎 The entity, Catena Trust Bank, N.A., would be a limited-purpose uninsured trust company headquartered in New York: no branches, no deposits, no lending. Pure custody, conversion, and settlement for fiat, stablecoins, and digital assets under the GENIUS Act framework.

Think Anchorage Digital’s charter playbook, rebuilt for non-human clients. Agents get verifiable identities, deterministic spending policies, kill switches, and full audit trails. Operators keep a control plane. The bet is that enterprises won’t hand agents real budget authority without a regulated fiduciary on the other side.

THE TAKEAWAY ✈️

What’s next? 🤔 First and foremost, we must note that the 30-day public comment window closes around late June 2026. If you want to gauge how seriously Washington takes autonomous AI as an economic actor, watch this docket. Of course, charter approvals typically take 12-18 months, and approval is far from certain. But acceptance of the filing alone tells you something: regulators are engaging with the premise that AI agents need their own financial identity framework, not a retrofit of human-centric KYC. Looking ahead, the second-order question nobody’s asking yet is this: if Catena gets chartered, who’s liable when an agent makes a bad trade? The charter creates a fiduciary wrapper, but the case law doesn’t exist. The first dispute will matter more than the first transaction. Either way, agentic finance is the future 🤖

ICYMI: Google Universal Cart and the Universal Commerce Protocol are Google’s bid to become the operating system of e-commerce ⚙️🛍️ [what Google Universal Cart is all about, why it matters & how it can transform e-commerce forever + bonus deep dives into Google’s UCP, how it just won Agentic Commerce & why AmEx skipped the protocol race altogether inside]

🧠 What else I’m watching

• AI Banking Companion 📱 CommBank is testing a conversational AI interface, CommBank Companion, within its app to help customers manage spending, savings, and financial goals. The tool analyzes live data to provide insights on budgets, subscriptions, cash flow, and business projections. Currently in employee and select business banking trials, it aims to shift banking from customer navigation to proactive, relevant information delivery. ICYMI: ChatGPT’s new personal finance feature is OpenAI’s bid to kill the fintech app economy 🤖📲 [what it’s all about, which fintechs should worry the most, how it connects to OpenAI’s Super App Play & what’s next + bonus deep dives into Anthropic & how it’s becoming Wall Street’s OS, their SME OS play, and what AI startup founders should build in 2026]

• TD AI Mortgage Boost 🏠 TD Bank’s new agentic AI model, developed at its Layer 6 R&D center, automates mortgage and home equity loan pre-adjudication, slashing processing time from 15 hours to under 3 minutes. The AI classifies documents, extracts data, validates income, and flags discrepancies before generating underwriter summaries - improving both accuracy and speed. This marks a foundational step in TD’s AI strategy, aiming to generate $1 billion in annual value by combining human expertise with AI efficiency. ICYMI:

• BNP-Mistral AI Defense 🛡️ BNP Paribas has renewed its three-year partnership with Mistral AI to bolster cyber defenses against vulnerabilities exposed by tools like Anthropic’s Mythos. With Mythos inaccessible to European banks, BNP is collaborating with Mistral to develop a similar cyber-focused AI model, addressing the speed and scale of emerging threats. The partnership also aims to simplify internal knowledge access, deploy an AI employee assistant, and create on-premises AI tools for sensitive processes like KYC. BNP emphasizes a multi-model approach, prioritizing performance, data sensitivity, and geographic alignment to drive operational and regulatory impact. ICYMI:

💸 Following the Money

• Scapia, an Indian outfit that combines co-branded credit cards with travel booking, has raised $63M in a funding round led by General Catalyst.

• Relay, a business banking and money management service for SMEs, has secured $50M in growth investment from General Catalyst.

• MoonPay announced its fourth acquisition of the year with Decent, as it launches MoonPay Trade.

👋 That’s it for today! Thank you for reading, and have a relaxing Sunday! And if you enjoyed this newsletter, invite your friends and colleagues to sign up:

Subscribe now

Subscribe now

On May 28, Anthropic announced a $65 billion Series H round at a whopping $965 billion post-money valuation, making it the most valuable private (AI) company on Earth. The same day, it released Claude Opus 4.8.

That timing is definitely intentional here. Opus 4.8 is the model Anthropic is betting its near-trillion-dollar story on.

Anthropic describes it as an upgrade on Opus 4.7 “with improvements across benchmarks” and “a more effective collaborator,” available at the same price ($5/$25 per million input/output tokens). The framing is clear as Opus 4.8 builds on 4.7’s foundation rather than breaking from it.

The prompting instincts you developed with 4.7 still apply. But there are meaningful new capabilities, and one genuinely new paradigm, that change what’s achievable if you know how to use them.

The biggest one is undoubtedly Dynamic Workflows in Claude Code: the ability for Claude to write its own orchestration scripts, spin up parallel subagents, and scale test-time compute trivially. Anthropic engineers have been using this as a daily driver for months. The rest of the world found out yesterday.

To make this as easy and seamless as possible, we’re also sharing Opus 4.8 Prompt Optimizer, a custom Claude skill that rewrites any prompt using the framework below automatically. Step-by-step setup instructions are at the end.

If you’ve read the Claude Opus 4.7 Prompting Playbook, the good news is that your existing prompts still work. But the new capabilities reward a specific set of upgrades, particularly around the effort stack, honesty directives, and Dynamic Workflows. This guide covers all of it.

Below is everything you need to get the best output this model can produce, every single time.

Step 0 - Choose the Right Model

Before writing a single word of your prompt, choose the right model. Anthropic’s current lineup includes Claude Opus 4.8, Claude Sonnet 4.6, and Claude Haiku 4.5, each optimized for different workloads.

→ Claude Opus 4.8 is now the flagship. It scores 69.2% on SWE-bench Pro, 88.6% on SWE-bench Verified, and 1890 Elo on GDPval-AA (121 Elo points ahead of GPT-5.5 on economically valuable professional tasks). It excels at long-horizon autonomous tasks, complex reasoning, and knowledge work. It costs more and runs slower via the API ($5/MTok input, $25/MTok output), but when the task demands genuine thinking - strategic analysis, multi-step research, complex autonomous execution - nothing else in the generally available lineup comes close.

→ Claude Sonnet 4.6 is still the balanced workhorse: strong reasoning at faster speed and lower cost. For most everyday tasks, Sonnet covers roughly 80% of use cases well.

→ Claude Haiku 4.5 remains the speed specialist: fastest, cheapest, ideal for high-volume, straightforward tasks like classification, extraction, and summarization.

💡 The rule of thumb: Start with Sonnet. Move to Opus when you need genuine cognitive depth, or when you’re running a Dynamic Workflow that demands long-horizon coherence. Drop to Haiku when speed matters more than intelligence.

The Framework

Anthropic’s 31-page+ documentation describes a hierarchy of techniques ordered by impact. Most people jump straight to advanced techniques and skip the fundamentals. That’s backwards.

Here’s the framework in the order that matters, updated for the specific behaviors of Opus 4.8.

1. Set the Effort Level (Most Important)

This is the single most important variable in your prompt, and most people never set it at all.

The effort parameter controls how much intelligence the model applies to a task. Like Opus 4.7, Opus 4.8 respects effort levels strictly, especially at the low end. The system remains: low, medium, high (default), xhigh, and max.

One change from 4.7: Opus 4.8 defaults to high effort (4.7 defaulted to xhigh), which Anthropic says spends similar tokens on coding tasks while performing better. Adaptive thinking is off by default and must be explicitly enabled.

The Effort Level System for Opus 4.8:

In the API:

client.messages.create(
    model="claude-opus-4-8",
    max_tokens=64000,
    thinking={"type": "adaptive"},
    output_config={"effort": "xhigh"},
    messages=[{"role": "user", "content": "..."}],
)

When running at max or xhigh, set your max_tokens to at least 64k - the model needs room to think, reason, and execute across tool calls and subagents.

Key Opus 4.8 update from the System Card: At minimum effort, Opus 4.8 matches the peak performance of Opus 4.7 at maximum effort on SWE-bench Pro. The floor has risen. But the ceiling has expanded too - xhigh and max now unlock capabilities (particularly in long-horizon agentic work) that Opus 4.7 couldn’t reach at any effort level.

For Dynamic Workflows specifically, enable ultracode mode: this uses workflows aggressively without requiring you to ask for them every time.

💡 Practical rule: If you see shallow reasoning on a complex task, raise the effort level before changing your prompt. That’s almost always the correct fix.

2. Be Specific - More Than You Think

The single highest-leverage prompting move is specificity. Opus 4.8, like 4.7 before it, is extremely literal. Vague prompts get scoped, not generalized. If you want an instruction applied across all sections, not just the first, you have to say so. If you want exceptional output, you have to request it explicitly.

Anthropic’s own recommendation still holds: think of your prompt as instructions to a brilliant but literal new hire on their first day. They’ll do exactly what you say, so say exactly what you mean.

❌ Weak prompt:

Write about market positioning.

✅ Strong prompt:

Analyze the 3 most effective market positioning strategies for B2B fintech companies targeting mid-market in a crowded category. For each strategy, explain what's driving its effectiveness, provide one specific company example, and assess whether it's likely to strengthen or weaken over the next 18 months. Apply this framework to all three strategies, not just the first.

The difference isn’t more words. It’s more specificity, explicit scope, and a format directive at the end.

3. Use XML Tags for Structure

This is Claude’s structural superpower, and still almost nobody uses it correctly.

Claude was specifically trained to recognize XML tags as structural markers. When your prompt has multiple components - context, instructions, data, constraints, output format - XML tags prevent Claude from mixing them up. Here’s what that looks like for a strategic decision:

<context>
You are helping me evaluate a potential Series A investment.
The company is a vertical fintech targeting logistics operators,
currently at $2.4M ARR growing 15% MoM.
</context>

<instructions>
Analyze the three key risks that most commonly derail vertical fintech companies
at this stage.
For each risk, explain the warning signs and what a founder should be doing
to mitigate them.
Apply this analysis to all three risks, not just the most obvious one.
</instructions>

<constraints>
- Be direct. Give me your honest assessment, not a balanced "it depends."
- Use specific examples from real companies where possible.
- Flag any assumptions you're making.
- Maximum 600 words.
</constraints>

Claude sees the tags and immediately understands that <context> is background (not a task to respond to), <instructions> is the actual task, and <constraints> are the guardrails.

Tag names are flexible - there’s no magic set of “correct” tags. Use whatever makes semantic sense: <background>, <rules>, <examples>, <output_format>. Consistency across your prompts matters more than the specific names you choose.

4. Show Claude What Good Looks Like

If one technique consistently separates good outputs from great ones, it’s this: show Claude what good looks like.

Instead of describing the tone, format, or style you want in abstract terms, provide two to three concrete examples. Claude will pattern-match against these far more reliably than it will follow descriptive instructions alone. Wrap examples in <example> tags (multiple examples in <examples> tags) so Claude distinguishes them from instructions.

<examples>
  <example>
    Input: "We need to cut 20% of the engineering budget"
    Output: "Reducing engineering spend by 20% requires prioritization
    across three areas: contractor headcount, infrastructure costs, and
    tooling licenses. Here's a phased approach that preserves our two
    highest-impact product initiatives..."
  </example>
</examples>

Now analyze this situation using the same approach:
"We need to extend our runway by 6 months without reducing headcount"

Anthropic recommends 3–5 examples for best results. You can also ask Claude to evaluate your examples for relevance and diversity, or to generate additional ones based on your initial set.

5. Prompt for Visible Reasoning (Chain-of-Thought)

For complex problems requiring analysis, multi-step reasoning, or strategic judgment, telling Claude to work through its reasoning before producing a final answer dramatically improves accuracy.

The simplest version: add “Think through this step by step before giving your final answer” to your prompt.

The more structured version uses tags to separate reasoning from output:

<instructions>
Evaluate whether we should expand into the Southeast Asian market this year.
Before giving your recommendation, work through the analysis inside
<analysis> tags.
Consider: market size and growth trajectory, regulatory requirements by
country, competitive landscape, our current operational capacity,
and capital requirements vs. expected payback period.
Then provide your final recommendation with a clear resource allocation
suggestion.
</instructions>

Forcing visible reasoning prevents Claude from pattern-matching to the most likely answer and back-filling justification after the fact.

Adaptive thinking in Opus 4.8: Claude dynamically decides when and how to think based on the effort setting and task complexity. At high and xhigh effort, deep reasoning is largely automatic for demanding tasks. Extended thinking with a fixed budget_tokens is no longer supported - adaptive thinking is the only thinking-on mode, and Anthropic’s evaluations show it reliably outperforms the old fixed-budget approach.

To guide reasoning explicitly, add this to your system prompt:

✅ “After receiving results, carefully reflect on their quality and determine optimal next steps before proceeding. Use your thinking to plan and iterate based on this new information, then take the best next action.”

6. Load Rich Context

Claude can only work with what you give it. The more relevant context you include, the more tailored and accurate the output becomes.

Upload documents. Paste data. Provide company background. Share your goals. Explain your audience. Don’t make Claude guess what you already know.

<background>
Our company builds financial infrastructure for neobanks in emerging markets.
We're Series B, $12M ARR, primarily serving West Africa and Southeast Asia.
Our main competitors are Banking-as-a-Service players like Railsbank, Synapse,
and local incumbents.
We differentiate on compliance coverage and local payment rail integrations.
</background>

<data>
[Paste your Q1 metrics, customer feedback, churn data, or whatever's relevant]
</data>

<task>
Based on this context, identify our three biggest growth opportunities
for the next quarter.
</task>

For long documents (20k+ tokens), put the documents at the top of your prompt, above your instructions and query. Queries at the end can improve response quality by up to 30% on complex, multi-document inputs.

7. Specify the Output Format

Don’t leave the structure of Claude’s response to chance. If you want a table, ask for a table. If you want a specific word count, state it. If you want an executive brief with defined sections, describe each section.

<output_format>
Respond with:
1. A one-paragraph executive summary (3-4 sentences max)
2. A comparison table with columns: Factor | Current State | Target State | Gap
3. A "Recommended Actions" section with 3 specific next steps, ranked by impact
</output_format>

Explicit format specs eliminate the most common frustration people have with AI: getting a 2,000-word essay when you wanted a concise brief, or getting bullet points when you needed flowing analysis.

8. Define Constraints - Especially “Don’ts”

Telling Claude what not to do is just as important as telling it what to do. Without constraints, Claude defaults to its training patterns, which can mean hedge-heavy output that sounds like a committee wrote it.

<constraints>
- Do NOT open with "In today's rapidly evolving landscape" or any variant
- Skip the preamble. Start with the most important insight.
- No bullet points — write in prose paragraphs
- If you're uncertain about a claim, flag it explicitly rather than hedging
  everything
- Maximum 500 words
- Be direct. I want your honest assessment, not a balanced "it depends."
</constraints>

One note specific to Opus 4.8 from the System Card: The model has a tendency toward over-elaborate refusals on certain categories of requests. If you’re getting unnecessary hedging or qualification on legitimate business tasks, add a clear statement of legitimate purpose to your prompt - e.g., “This analysis is for an internal investment committee memo. Please provide a direct, unqualified assessment.”

9. Control Verbosity

Opus 4.8 calibrates response length to how complex it judges the task to be - short on simple lookups, much longer on open-ended analysis. If your use case requires specific verbosity, tune it explicitly.

To decrease verbosity:

“Provide concise, focused responses. Skip non-essential context and keep examples minimal.”

To increase depth:

“This is for board-level review. Go deep on every dimension - don’t compress the analysis.”

Key insight from Anthropic itself: positive examples showing appropriate concision work better than negative instructions like “don’t be verbose.” Show Claude a response at the length and depth you want, and it will match that pattern far more reliably than it will follow abstract length instructions.

10. Use Dynamic Workflows (The New Paradigm)

This is new to the Opus 4.8 era, and it’s the most powerful capability yet.

What it is: Dynamic Workflows in Claude Code allow Claude to write its own orchestration scripts that spin up tens to hundreds of parallel subagents, instead of making tool calls one at a time. The control flow is code, which means Claude won’t drift or forget halfway through thousands of files. The harness is encoded.

How it works: Mention the word “workflow” in any Claude Code prompt, and Claude will spin one up automatically. For aggressive workflow use without prompting, enable effort ultracode in your settings.

What it unlocks:

• Massively parallel execution: A five-agent team can handle BrowseComp tasks using just 20% of the latency of a single agent with a 10M-token limit, while scoring higher. The Orchestrator with Blocking Subagents configuration reached the highest scores overall.

• Adversarial verification: Route critical work through adversarial judges that each attack the output from a different lens. You can go as far as a tournament-style bracket, scoring competing plans against each other.

• Scalable test-time compute: Workflows let you scale test-time compute trivially. There has never been an easier way to do this in Claude Code.

• Reusable automation: Because workflows are scripts under the hood, you can save them, commit them to your repo, share them with your team, and turn them into skills.

The flagship real-world example: Jarred Sumner (creator of Bun) used Dynamic Workflows to rewrite the entire Bun runtime from Zig to Rust, roughly 750,000 lines of code, eleven days from first commit to merge, 99.8% of the existing test suite passing. The workflow mapped Rust lifetimes for every struct field in the Zig codebase, then wrote every .rs file as a behavior-identical port of its .zig counterpart, hundreds of agents working in parallel with two reviewers on each file. A fix loop then drove the build and test suite until both ran clean.

When to reach for a workflow:

→ The task is too big for a normal session (codebase-wide bug hunt, large refactor, multi-source research).

→ You need higher confidence that something is right (critical features, financial models, security reviews).

→ You want to parallelize work across many files or domains simultaneously.

🔥 Cost warning: Workflows can get expensive due to all the parallel agents. Get a feel for token usage on smaller tasks before unleashing them on larger ones. You can disable them with config or enterprise settings if needed.

Best practices for Dynamic Workflows:

1. Trigger intentionally. Say “workflow” explicitly for important tasks. For routine work, keep workflows off to manage costs.

2. Enable ultracode for coding sprints. When you’re in a focused engineering session, effort ultracode removes the friction of triggering workflows manually.

3. Use adversarial judges for critical outputs. Don’t just generate - verify. Route the output through a judge agent with explicit attack criteria.

4. Save your best workflows as skills. If a workflow solves a recurring problem well, commit it. This is institutional knowledge that compounds.

5. Set a token budget before large tasks. Ask Claude to estimate token usage before running a large workflow, then decide if the cost is justified.

6. Use blocking subagents for accuracy-critical tasks. Async subagents reduce latency but the Orchestrator with Blocking Subagents configuration scores highest when accuracy matters more than speed.

7. Let Claude choose the architecture. Don’t over-specify the workflow structure. Claude knows when to parallelize, when to serialize, and when to use a tournament bracket.

8. Test on a sample first. On large codebases or research tasks, run the workflow on a representative 10% before committing to the full task.

Putting It All Together

Here’s what a properly structured Claude Opus 4.8 prompt looks like when you combine all the techniques covered above:

<context>
I'm the CEO of a B2B fintech startup ($8M ARR, 45 employees).
We're deciding whether to raise a Series B now or extend runway and raise
in 18 months.
Current runway: 14 months. Revenue growth: 12% MoM. CAC payback: 8 months.
</context>

<instructions>
Analyze both timing options. Before giving your recommendation, work through
the trade-offs in <analysis> tags, considering:
- Current market conditions for fintech Series B rounds
- Our specific metrics relative to typical Series B benchmarks
- The risk/reward of raising now vs. in 18 months at potentially better metrics
- What we should use the 18 months to optimize if we extend
Then provide a clear recommendation with a specific action plan.
Apply your analysis to both options equally — don't weight one by default.
</instructions>

<constraints>
- Be direct. Give me your honest read, not a balanced "it depends."
- Use specific benchmarks from comparable fintech Series B raises where possible.
- Flag any assumptions you're making about market conditions.
- Keep the total response under 700 words.
</constraints>

<output_format>
1. Analysis (in <analysis> tags)
2. Recommendation (2-3 sentences, clear and direct)
3. 90-day action plan (5 specific actions, whether we raise now or extend)
</output_format>

This prompt is clear, structured, specific, and constrained. It tells Claude exactly what to do, how to think about it, what to avoid, and how to format the response. The output from this kind of prompt is categorically different from “Should I raise a Series B now or wait?”

10 Mega Prompts Engineered for Claude Opus 4.8

As always, understanding the principles is one thing. Having them already engineered into prompts you can use today is another.

Below are 10 Mega Prompts, each structured using every technique above and calibrated for Claude Opus 4.8’s specific behaviors. Each covers a high-leverage task for founders, operators, and investors.

How to use these prompts:

• Fill in every [PLACEHOLDER] before submitting. Blank placeholders produce generic output.

• Enable web search for prompts that reference real-time data: Research (#1), Competitor Analysis (#8), Stock/Deal Analysis (#9), and Reverse-Engineering (#10). Without it, Claude will rely on training data and flag confidence gaps.

• Use xhigh or high effort for all 10 prompts. For #9 and #10, start at xhigh or max.

• Mix and match: you can add chain-of-thought <analysis> tags, output constraints, or anti-patterns from one prompt into any other.

• Alternatively, you can use Opus 4.8 Prompt Optimizer skill that rewrites any prompt using this framework automatically. Step-by-step setup instructions are at the end.

Prompt #1 - Deep Research Brief

Produces an executive brief with sourced findings, expert perspectives, and contrarian views.

Read more

Subscribe now

Small businesses have always had the ideas. They’ve rarely had the capacity.

And yet they carry the economy. There are 36 million small businesses in the U.S. alone - 99.9% of all firms - contributing 43.5% of GDP and employing 46.5% of the private-sector workforce. Globally, SMEs represent roughly 90% of all businesses and generate up to 70% of both employment and GDP across OECD countries. They are not a footnote to the economy. They are the economy.

Yet for decades, the tools that could change how they operate - real automation, intelligent data analysis, multi-system orchestration - were priced and engineered for enterprises with dedicated IT teams and six-figure software budgets. The owner who handles sales, operations, invoicing, marketing, HR, and customer service simultaneously has never had access to the intelligent support that enterprise teams take for granted.

That gap is exactly what Claude for Small Business is built to close.

On May 13, 2026, Anthropic launched it as a plugin for Claude Cowork that puts Claude inside the tools your business already runs on: QuickBooks, PayPal, HubSpot, Canva, DocuSign, Slack, Google Workspace, and Microsoft 365.

One toggle. No developer. No API keys. No third-party integration platform required.

What this means in practice: Claude chases your overdue invoices, drafts tone-matched follow-ups based on each client’s payment history, and queues them in Gmail for your approval - all from a single command. It runs a Monday morning business pulse that pulls your cash position, pipeline, and calendar into a one-page brief. It builds a hiring packet - job post, interview guide, offer letter - and routes it through DocuSign. It plans your payroll, flags cash shortfalls, creates Canva assets in your brand colours, and reviews incoming contracts for liability traps.

It does all of this without sending, posting, or paying a single thing without your explicit approval.

The economics are honestly wild: at $20/month for Claude Pro, even a conservative estimate of five recovered hours per week at $50/hour owner-time valuation returns $250 per week against a $20 subscription. The ROI is not marginal at all, but it depends on the setup time you invest upfront.

This guide was built for this exact reason and covers every single step:

→ What Claude for Small Business actually is

→ How to set it up correctly

→ How to build the CLAUDE.md business brain that makes everything smarter

→ The most powerful workflows with copy-paste prompts,

→ A deep dive into every integration, and

→ The real limitations and risks you need to understand before letting Claude anywhere near your finances.

Let’s dive in.

The Architecture: Three Layers You Need to Understand

Understanding the structure prevents most of the confusion around what Claude for Small Business can actually do.

Read more

Good morning & happy Wednesday! Today we’re diving deep into another British digital banking gem Starling Bank, which just published its 2026 annual report (unpacking the most important financial facts & figures, how it stacks against Monzo & Revolut, & why Starling’s balance sheet is a fortress + bonus deep dives into the latest financials of Monzo & Revolut inside), and why your next broker and investment advisor could be Claude AI or ChatGPT (what Liquid Co-Invest is all about, why it’s interesting & what it tells us about the future of agentic finance + bonus dive into how this perfectly fits the AI products Anthropic itself told founders to build in 2026 & the full Claude Playbook for Finance inside). So let’s jump straight into the good stuff 🌶️

Starling Bank’s 2026 financials: 2.5x Monzo’s profit on half the revenue, and a £70M SaaS pipeline that changes the IPO math 📈🧠

Following the money 💸 All three biggest UK neobanks finally published their annual results, with two of them within days of each other. The headlines wrote themselves: Revolut posted £1.7 billion in pre-tax profit on £4.5 billion in revenue. Monzo crossed £1.7 billion in revenue and added 3 million customers. Starling’s revenue fell 5.6% 🤕

Starling clearly lost the headline war. But it may have won everything else.

On £887M in revenue, Starling produced £217M in statutory profit, a 24.5% PBT margin. Monzo, on nearly double the revenue, managed £87M, a margin of roughly 5%. Revolut is in a different category entirely: a global fee-income platform serving now over 70 million customers across 39 countries, barely lending at all. Starling is thus the only one of the three that looks like an actual bank, lends like an actual bank, and earns like the best of them 🏦

But the number that should change how you think about Starling isn’t in the P&L. It’s buried in the platform business.

Engine, Starling’s SaaS banking-as-a-service arm, generated just £10.9M in revenue last year. Rounding error. Except it now has £70M in committed annual recurring revenue, four live clients across five countries, and a single partnership (Tangerine, a Scotiabank subsidiary) that alone will migrate over 2 million accounts onto the platform over the next decade. If Engine scales to £100M in ARR, a 10x revenue multiple on that segment alone could add £1B or more to Starling’s IPO valuation, nearly matching its entire equity book value today. Not too shabby!

That makes Starling two businesses stitched together inside one balance sheet: a high-margin UK bank with a £525M capital surplus, and a nascent fintech infrastructure company with global reach and zero balance-sheet consumption. The market will eventually have to price both. It hasn’t done so yet.

So let’s dive deeper here and unpack the full financial picture: the Monzo and Revolut comparisons in detail, why Starling’s balance sheet is a fortress (and why Monzo’s lending book tells a very different credit story), how Engine’s pipeline and committed ARR stack up against the valuation math, what the FCA restrictions still in place mean for the growth re-acceleration thesis, and the specific entry price at which the risk/reward turns decisively in Starling’s favour ahead of a likely 2027–2028 IPO.

Read more

Good morning & happy Tuesday! Today we’re diving deep into British neobanking gem Monzo, which just published its 2026 annual report (breaking down the unit economics that matter, the Revolut comparison investors keep getting wrong, the governance drama, IPO chances & what’s next for Monzo + bonus deep dive into Revolut’s latest financials & their foundational AI model for money inside), and Catena Labs that wants a bank charter for AI Agents (what they are trying to do, why it’s worth paying attention & how Catena stacks agents other players in agentic finance space + bonus deep dive into Google’s bid to become the OS for all e-commerce, Circle’s Agent Stack play, and a full guide on how to build an AI Agent from sractch inside). So let’s jump straight into the interesting stuff 🌶️

Monzo’s 2026 financials: the £1.7 billion revenue machine that still has to prove it can escape the UK’s Orbit 🏦🇬🇧

Following the money 💸 Monzo Bank makes more money per customer than Revolut. That sentence would have been unthinkable two years ago. But recently dropped 2026 annual report confirmed it: Monzo now generates £167 in revenue per active personal customer, while Revolut generates £66. The gap is the difference between a primary bank 🏦 and a multi-currency spending card 💳, and it’s the number that will anchor every IPO conversation Monzo has this year.

The headline figures are definitely strong.

• Revenue surged 39% year-on-year to £1.71B.

• Gross profit broke £1B for the first time.

• Statutory pre-tax profit hit £87.3M (up 44%), or £172.6M adjusted for an FCA fine and restructuring charges. Three consecutive years in the black.

• A customer base of 15.2M, with one in five UK adults now banking with Monzo.

• Deposits up 55% to £25.7B. Four revenue streams each clearing £300M, up from one a year ago.

By every growth metric, this is a neobank that has crossed into proper banking territory.

Yet, the investment case isn’t about what Monzo has done. It’s about what happens next, and that’s where the picture gets complicated.

The board pushed out its previous CEO partly over frustration with the company’s UK concentration. His replacement, Diana Layfield, exited the US within weeks of taking charge and is betting the company’s next phase on European markets where Revolut already has a decade-long head start and millions of entrenched customers. Monzo’s statutory profit margin is 5.1%. Revolut’s is 38%. The CET1 ratio fell from 56% to 33% in a single year. And Layfield has told the FT she’s “not in a hurry” to IPO, which is either confidence or a warning that the timeline is longer than the market expects.

So let’s dive deeper and break down the unit economics that matter, the Revolut comparison investors keep getting wrong, the governance drama behind the CEO transition, and where the valuation range of £6-10 billion starts to look either like a bargain or a stretch. The answer depends on a single question: can Monzo replicate in Dublin and Barcelona what it built organically in London?

The Numbers That Matter

Read more

Subscribe now

Thanks to AI, finance teams are now shipping board decks, closing books, and building models under the same pressure as always - except now with fewer people. Claude AI can cut a 13-week cash forecast from four hours to 40 minutes, but only if you set it up right. A bad prompt still produces a confident, wrong DCF. A good one produces an auditable first draft that your controller can actually use.

This guide gives you the exact setup: a 60-minute configuration walkthrough, seven copy-paste prompts (DCF, LBO, three-statement model, 13-week cash forecast, unit economics, credit analysis, and monthly close), and the specific failure modes that will burn you if you don’t know about them. Every prompt uses XML-structured instructions optimized for Opus 4.7’s literal instruction-following - the formatting pattern that consistently produces the tightest financial outputs.

Whether you’re a CFO running a $30M SaaS company, an associate building an IC deck, or an FP&A lead trying to close in five days instead of ten, the prompts below are ready to drop into your next session.

And once you’ve built these workflows, the Cowork Automation Playbook and the /goal Guide for AI Agents show you how to put them on autopilot - scheduled deal scoring, autonomous research runs, morning briefings that generate themselves.

Why Claude Opus 4.7 for Finance

Read more

If you’re not a subscriber, here’s what you missed this week:

1. SpaceX IPO: Why the Biggest Listing in History Might Be Uninvestable at $1.75 Trillion 🪐 [The S-1 reveals a $7 billion Starlink profit machine subsidizing $30B in annual AI cash burn, while Elon Musk holds 85% of the vote with no sunset, no independent board, and no path to remove him]

2. Ex-OpenAI’s Aschenbrenner Bet $5.5B on AI Chips. Now He’s Betting $8.5B Against Them 💸 [Leopold Aschenbrenner’s AI hedge fund Situational Awareness went from pure AI long to 62% short semiconductors in one quarter. Here’s what changed and why the thesis still holds]

3. Coatue’s May 2026 Report: The $12 Trillion AI Bet and Who Ends Up on the Wrong Side 📊 [Coatue’s latest public markets report reveals the most extreme winner/loser split in tech stock history, a $12 trillion AI capex wave, and one brutal framework that predicts exactly who wins]

4. Anthropic Just Told AI Founders Exactly What to Build in 2026 🦄 [1 million conversations. 9 consumer AI domains. A full founder playbook - plus where Anthropic’s own products will and won’t compete]

5. ChatGPT’s new personal finance feature is OpenAI’s bid to kill the fintech app economy 🤖📲 [what it’s all about, which fintechs should worry the most, how it connects to OpenAI’s Super App Play & what’s next + bonus deep dives into Anthropic & how it’s becoming Wall Street’s OS, their SME OS play, and what AI startup founders should build in 2026]

6. Google Universal Cart and the Universal Commerce Protocol are Google’s bid to become the operating system of e-commerce ⚙️🛍️ [what Google Universal Cart is all about, why it matters & how it can transform e-commerce forever + bonus deep dives into Google’s UCP, how it just won Agentic Commerce & why AmEx skipped the protocol race altogether inside]

7. The Complete Claude /goal Guide for AI Agents 🤖 [What Claude /goal actually is, why most invocations fail, and how to write production-grade goal conditions for financial research, backtesting, and live market monitoring]

8. How to Build an AI Agent from Scratch (With Working Code) 🤖 [The design framework, working code, and hard-won lessons that take you from “I want to build an AI Agent” to a working one - in under 60 minutes]

9. Claude Cowork Commands, Scheduled Tasks & Automation Workflows: The Operator’s Playbook 🤖 [Every real slash command, the exact prompts that run deal flow and investor updates on autopilot, and the failure modes that will cost you hours if you don’t know them first]

10. The System for Never Hitting Claude’s Limits 🤖 [Most users burn the majority of their token allocation on architecture mistakes, not actual work. Here’s the operational framework that fixes it]

As for today, here are the 2 incredible FinTech stories that are transforming the world of financial technology as we know it. This was yet another insane week in the financial technology space, so make sure to check all the above stories.

Revolut is rebuilding its revenue mix before anyone sees the S-1 🤑🏦

Following the money 💸 FinTech giant and Europe’s most valuable private company Revolut, doesn’t have a banking problem. It has a revenue-quality problem, and two moves in the past week show it knows exactly how to fix it.

Let’s take a look.

More on this 👉 The London fintech is launching a private banking arm targeting clients with £500,000 or more in deposits, with services rolling out later this year. The FCA just granted its trading subsidiary permission to offer leveraged products, managed portfolios, and investment advisory.

Meanwhile, a leaked internal memo from CEO Nik Storonsky on May 15 designated Revolut Business as the company’s top strategic priority, backing it with dedicated growth teams, a £1,000-per-head employee referral bonus for new business accounts, and a commitment to bundle business banking into every new market starting 2027. Not too shabby!

Zoom out 🔎 Read these two developments together. Revolut posted £4.5 billion in revenue last year and £1.7 billion in pre-tax profit. Impressive numbers, but 76% of revenue comes from fees on cards, FX, subscriptions, and crypto trading. Public markets will discount that mix. It’s volatile, competition-exposed, and hard to model forward. Private banking (AUM-based fees, advisory, lending spreads) and sticky B2B relationships (£940 estimated ARR per business customer, 800,000 accounts and climbing) are the antidote.

The private banking play targets a specific gap. Coutts now requires roughly £3 million minimums. Revolut already has millions of UK users in the £500,000-to-£3 million range who are barely monetized. Hiring ex-HSBC private bankers to pair with a digital-first platform is the right hybrid model for mass-affluent clients who want service without the theater.

We can note that wealth already generated £663 million last year. Even modest early traction, say 5,000 clients at £1.5 million average AUM, adds £5-8 billion in managed assets and tens of millions in high-margin recurring revenue.

THE TAKEAWAY ✈️

What’s next? 🤔 Looking ahead, there are several things to watch. First and foremost, the IPO isn’t coming before 2028, and the $200 billion valuation target demands a story that looks more like JPMorgan than Robinhood. Thus, every quarter between now and then (if/when possible), track two numbers: AUM growth in the new wealth segment and B2B’s share of gross profit (already approaching 20%). If both inflect upward, Revolut gets its multiple re-rating before it ever files. If the white-glove service falls flat or credit underwriting stumbles in new markets, the narrative unravels fast. All in all, we can definitely agree that the strategy is correct here. Execution is the only remaining question, and it’s not a small one. But so far, Revolut hasn’t had any problems executing…

ICYMI: Inside Revolut’s PRAGMA: The Foundation Model Trained on 40 Billion Banking Events 🧠 [Architecture, performance benchmarks vs. Stripe, Mastercard, and Visa, regulatory risks, and why PRAGMA may be the most consequential AI bet in consumer finance]

Revolut is the most profitable FinTech on Earth, and it hasn’t even started lending yet 🤯📈 [unpacking the fee-income fortress that makes Revolut structurally different from every competitor, the lending optionality coiled inside that balance sheet, the regulatory dominoes now falling in sequence, and whether $75B is expensive or whether the market is about to find out it was way too cheap + more bonus reads on Revolut, how it’s leveraging AI & bonus deep dives into the latest financials of Nubank, SoFi, Robinhood, and Coinbase inside]

OpenAI is paying for YC’s next batch in tokens it prints itself 🫣💸

The news 🗞️ OpenAI co-founder & CEO Sam Altman just found a way to turn inference compute into venture capital.

Let’s unpack this.

More on this 👉 OpenAI is now offering every startup in YC’s Spring 2026 batch - roughly 169 companies - $2 million in API credits through an uncapped SAFE. That’s ~$338 million at retail prices, paid in a currency OpenAI manufactures at a fraction of face value as marginal inference costs keep falling.

In return, OpenAI gets equity that converts at the next priced round, plus something worth more than any cap table line: deep visibility into what the sharpest AI-native founders are actually building, at cohort scale, in real time.

And that’s hard to put a price tag on.

Zoom out 🔎 Of course, the playbook isn’t new. Yuri Milner’s 2011 blanket bet on YC through SV Angel was the template - $150K cash per company, broad exposure, no picking. Altman, who was at YC when that deal went down, is running the same play with a critical twist: he’s paying in his own product. Every “dollar” of credits consumed flows back as usage data, training signal, and switching costs. The founders get transformative compute runway. OpenAI gets lock-in dressed up as patronage 😎

Altman’s framing - “tokenmaxxing” over headcount - tells you exactly where he wants these companies to land architecturally: inference-heavy, deeply coupled to OpenAI’s stack, expensive to migrate. For teams already planning to burn through trillions of tokens, the dilution math on an uncapped SAFE looks tolerable.

For those building thin orchestration layers on top of foundation models, taking this deal is handing your playbook to the one company best positioned to replicate it inside ChatGPT.

THE TAKEAWAY ✈️

What’s next? 🤔 Looking ahead, here’s what you should watch. First, competitive response: if Anthropic or Google counter with credits-for-equity programs (or credits with no equity attached), the subsidy war gets interesting fast. Second, post-credit behavior: when the $2M runs dry, do these startups stay on OpenAI or scatter to multi-model architectures? That answer will tell you whether this was customer acquisition or ecosystem capture. Third, product overlap: track OpenAI’s feature launches over the next 18 months against what gains traction in this batch. If the pattern rhymes with Amazon Marketplace or Apple’s App Store history, founders will notice, and the next cohort won’t sign as easily.

ICYMI: ChatGPT’s new personal finance feature is OpenAI’s bid to kill the fintech app economy 🤖📲 [what it’s all about, which fintechs should worry the most, how it connects to OpenAI’s Super App Play & what’s next + bonus deep dives into Anthropic & how it’s becoming Wall Street’s OS, their SME OS play, and what AI startup founders should build in 2026]

🧠 What else I’m watching

• More Bank AI Layoffs 🏦 Standard Chartered plans to cut 7,800 back-office jobs - 15% of its 52,000 support roles - over four years, targeting Chennai, Bengaluru, Kuala Lumpur, and Warsaw. CEO Bill Winters frames the move as replacing lower-value tasks with AI-driven automation, not cost-cutting, while some staff will be reskilled. The bank is investing in AI to streamline operations, positioning the technology as a key enabler for efficiency and transformation in its workforce strategy. Yet another reminder to productize yourself 😉 ICYMI:

• Fintech AI Skills Drive 📈 Zopa and ClearScore launched Jobs2030, a coalition upskilling 100,000 fintech and banking professionals in AI by 2030, with 22 members already on board. The free initiative, built around technology, content, training, hiring, and advocacy, offers five courses and 12 modules by year-end, tailored to compliance, engineering, operations, and product roles. Training on Zenarate integrates with LLMs like Gemini and ChatGPT, combining AI coaching, simulations, and real-world scenarios. Backed by Innovate Finance, The City of London Corporation, and venture firms, the program aims to deliver practical AI skills and foster better customer outcomes through industry-led expertise. ICYMI:

• AI Streamlines Vendor Risks 🤖 Deutsche Bank’s TPRM AI accelerates third-party vendor risk assessments, cutting document review time from 30 minutes to under two minutes for up to five files. Developed in-house, the system uses sequential AI agents to retrieve control questions, draft responses, and propose outcomes with 90% accuracy, followed by human validation. This cloud-based solution scales with demand, reducing onboarding timelines and costs while maintaining regulatory compliance. ICYMI:

💸 Following the Money

• Mercury, a provider of banking services for entrepreneurs and the tech industry, has raised $200M in a Series D at a $5.2B valuation.

• Stablecoin-powered digital banking and investment platform Fasset has raised $51M in Series B funding.

• National Australia Bank has acquired account-to-account payments platform Banked. Financial terms were not disclosed.

👋 That’s it for today! Thank you for reading, and have a relaxing Sunday! And if you enjoyed this newsletter, invite your friends and colleagues to sign up:

Subscribe now

Subscribe now

SpaceX dropped a 200,000-word S-1 on Wednesday evening. Buried inside is the financial blueprint for the largest IPO ever attempted - roughly three times the size of Saudi Aramco’s record-setting 2019 offering - at a valuation that would make SpaceX worth more than every publicly traded company on Earth except a handful of tech giants. The asking price: a whopping $1.75 trillion 🤯

The numbers tell a story of two companies forced into one balance sheet.

→ Starlink, the satellite broadband business, generated $7.2 billion in EBITDA last year on 63% margins, doubled its subscriber base to over 10 million, and operates 75% of all maneuverable satellites in orbit. It is, by almost any measure, the most valuable connectivity asset built in a generation 🛰️

→ Then there is xAI - Musk’s AI division, merged into SpaceX earlier this year - which burned $6.4 billion in operating losses and is spending at a $30 billion annualized capex rate with Grok trailing every frontier AI competitor. The AI segment alone is responsible for dragging a profitable company into an annual net loss of $4.9 billion. Outch 🤖

And then there is the governance. Musk will hold 85.1% of voting power after the IPO. He serves as CEO, CTO, and Chairman. His compensation is tied to milestones that include a Mars colony of one million inhabitants. He cannot be fired. Three of the largest pension funds in America - CalPERS, the NYC Comptroller, and NYS Comptroller, representing $1 trillion in combined assets - have already publicly demanded structural reforms. SpaceX rejected every one.

What’s interesting is that the S-1 also discloses a $45 billion computing deal with Anthropic that is cancelable on 90 days’ notice, a $60 billion option to acquire AI startup Cursor with a $10 billion breakup fee, $131 million in Cybertruck purchases from Tesla at retail price, and a $20 billion bridge loan that matures in 16 months.

The prospectus also runs to 37 pages of risk factors 😬 Oh, and the total addressable market SpaceX claims? $28.5 trillion, a figure that includes asteroid mining and Mars colonization 🪐

The key question now is not whether SpaceX has built something extraordinary. It clearly has. The question is whether, at 94x revenue, 266x EBITDA, and with a governance structure that gives public shareholders no voice and no recourse, the SpaceX IPO is worth your time and money, and whether its price leaves any margin of safety at all.

Our analysis of the full S-1 says it does not.

Valuation & Deal Structure: The Largest IPO in History

Read more

Good morning & happy Thursday! Today’s issue is especially interesting as we’re diving deep into Google, which is now going all in to become the operating system for the whole e-commerce (what Google Universal Cart is all about, why it matters & how it can transform e-commerce forever + bonus deep dives into Google’s UCP, how it just won Agentic Commerce & why AmEx skipped the protocol race altogether inside), OpenAI that just found a way to turn inference compute into venture capital (what their deal with Y Combinator is all about, what are the risks & challenges here + bonus deep dive into OpenAI, which is now trying to kill the FinTech app economy & how Anthropic told AI founders exactly what to build in 2026 inside). So let’s jump straight into the finnovative stuff 🌶️

Google Universal Cart and the Universal Commerce Protocol are Google’s bid to become the operating system of e-commerce ⚙️🛍️

The BIG News 🔥 Arguably one of the most consequential announcements at I/O 2026 wasn’t another AI model. It was a shopping cart.

Google is no longer content to sit at the top of the funnel. With Universal Cart and the Universal Commerce Protocol (UCP), it’s clearly making a play to own the entire commerce stack: the intent layer, the reasoning layer, and now the protocol layer that connects every merchant to every AI agent.

The pattern is unmistakable if you’ve watched Google before. Define the open standard. Get the ecosystem to adopt it. Then win on the implementation no one else can match.

Let’s dive deeper into this, break down how the product works, what UCP actually standardizes, who benefits, who gets squeezed, and the specific second-order effects investors and operators should be tracking over the next 12 to 24 months.

Read more

Subscribe now

Three weeks ago, I published a deep dive on Leopold Aschenbrenner and Situational Awareness LP, the $225M-to-$5.5B AGI-infrastructure hedge fund that quietly became the most intellectually coherent portfolio in public markets. The thesis, in one line, is this: buy the bottlenecks of the AGI buildout, short what AI eats.

Power, compute, memory, optics, fabs on the long side. Legacy IT services on the short side. Simple, sequenced, almost monastic.

In fact, it’s the same split Coatue’s May 2026 report frames as sellers of scarcity vs buyers of scarcity, and the two theses validate each other almost line for line.

Then, earlier this week, the latest 13F dropped. And the book broke its own template.

Gross notional jumped from $5.52 billion to $13.68 billion, a 148% expansion in 90 days. The book went from 29 to 37 disclosed positions across 29 distinct tickers (eight names are now held on both sides simultaneously).

But the headline isn’t the size. It’s the shape.

For the first time in the fund’s history, the disclosed book is net short: 38.1% long, 61.9% short. The short side is a wall of single-name puts on every Tier-1 name in the semiconductor stack: NVDA, AVGO, AMD, MU, TSM, ASML, INTC. Plus ORCL puts. Plus, SMH ETF puts back in size as the largest position in the book at 14.94%.

An important caveat before we go further: 13F filings disclose put and call notional but do not specify whether the fund is long or short those options. The near-universal read of this filing, and the one that fits the fund’s prior pattern of using puts as directional shorts, is that these are bought puts expressing a bearish view. That interpretation is what follows. But directional intent cannot be confirmed from the filing alone.

YTD return on the disclosed positions: +39.45%. SPY: +8.40%. That’s 31 percentage points of alpha in just 4.5 months 🤯

You can watch the whole book live on the Situational Awareness tracker, direction-corrected for P&L, bucket exposure, and the full overlay structure. It updates with every filing.

→ So does this break the thesis? No.

→ Does it change everything about how to express it? Yes.

And anyone running an AI-infrastructure book, building in the space, or trying to figure out what to do with the semis they bought in 2024 needs to update their priors immediately.

Let’s walk through it all.

What Actually Changed in Q1 2026

Read more

Good morning & happy Tuesday! Today we’re diving deep into OpenAI, whose new personal finance features inside ChatGPT is a bid to kill the FinTech app economy (what it’s all about, which fintechs should worry the most, how it connects to OpenAI’s Super App Play & what’s next + bonus deep dives into Anthropic & how it’s becoming Wall Street’s OS, their SME OS play, and what AI startup founders should build in 2026), and fintech giant Revolut that’s actively rebuilding its revenue mix before the IPO (two interesting moves from Europe’s most valuable startup, what they tell us & what’s next + bonus deep dives into Revolut’s latest 2025 financials & foundational AI model PRAGMA inside). Let’s jump straight into the incredible stuff 🌶️

ChatGPT’s new personal finance feature is OpenAI’s bid to kill the fintech app economy 🤖📲

The news 🗞️ The fastest way to destroy a software category is to make it a feature inside something everyone already uses.

OpenAI just connected ChatGPT to 12,000 banks and brokerages. Chase, Fidelity, Schwab, Robinhood , AmEx - all feeding real transaction data into a conversational AI layer powered by GPT-5.5. For the hundreds of millions of users already asking ChatGPT money questions, the answer now comes grounded in their actual spending, investments, and liabilities.

For every fintech startup that built an AI budgeting app on top of Plaid over the past three years, the threat is now existential. And for the broader AI industry, this is the clearest signal yet of where the Super App race is heading: whoever owns your financial context owns the most defensible consumer relationship in tech.

Let’s take a closer look at this, break down who this kills, who survives, what Anthropic’s opposite bet tells us about where AI finance is actually going, and the second-order regulatory and competitive dynamics most are missing.

Read more

Subscribe now

Coatue Management doesn’t publish market updates often. When they do, the industry listens.

The May 2026 edition is their most urgent yet. Despite tariff escalations, a Middle East oil war, 180-degree interest rate repricing in 12 weeks, and relentless AI bubble warnings, the Nasdaq just hit an all-time high. April 2026 was the best single month for equities since COVID’s emergency rebound in April 2020.

Markets are clearly not panicking. They are pricing in a structural reality that most people haven’t fully absorbed yet.​

OpenAI and Anthropic - two companies that barely existed as revenue businesses four years ago - have combined to reach $55B in annualized run-rate revenue, surpassing Salesforce, Adobe, ServiceNow, and Workday combined. Seagate stock has nearly 8x’d in a year. Micron’s operating margins went from 16% to 69%. Hyperscalers are on track to commit $700B+ in capital expenditures this year alone. And Coatue projects that between 2026 and 2031, $12 trillion in AI capex will be deployed - the largest coordinated capital mobilization in the history of private industry 🤯

But not everyone wins. In fact, right now, most companies are losing.

The core of Coatue’s framework is a single, devastatingly simple idea: the market is rewarding sellers of scarcity and punishing buyers of scarcity. Understand which side of that line your company, your portfolio, or your career sits on, and you’ll understand everything happening in markets right now.

Below is the full section-by-section breakdown of every signal in the Coatue report, cross-validated with live market data covering the macro backdrop, the anatomy of the winner/loser split, the agentic AI architecture shift, the $12T capex pipeline, the fracture of the Mag 7, Physical AI as the next mega-wave, and the most actionable takeaways for founders, operators, and investors in this market.

We’re also including two companion deep dives. First, how Leopold Aschenbrenner turned $225M into $5.5B in twelve months by betting on the AI infrastructure layer no one else was watching - the full portfolio breakdown, the 7-rule playbook, and the framework behind the best-performing hedge fund of the cycle. Second, Anthropic that accidentally handed AI founders the most precise product-market-fit signal ever published - 9 consumer domains, domain-by-domain competitive maps, and the exact line between where Anthropic will and won’t compete.

Let’s dive in.

The Macro Stage - Resilience Despite Everything

A Market That Shrugs Off Shocks

Read more

If you’re not a subscriber, here’s what you missed this week:

1. Anthropic Just Told AI Founders Exactly What to Build in 2026 🦄 [1 million conversations. 9 consumer AI domains. A full founder playbook - plus where Anthropic’s own products will and won’t compete]

2. Claude Opus 4.7: The Complete Prompting Playbook 🧠 [How to unlock the full intelligence of Anthropic’s most capable generally available model - with 10 ready-to-use Mega Prompts engineered for the way builders and investors actually work]

3. He Got Fired From OpenAI at 22. Then He Turned $225M Into $5.5B by Betting on the AI Infrastructure No One Else Was Watching 💸 [Inside Leopold Aschenbrenner’s Situational Awareness Fund: the AGI investment thesis, the full portfolio breakdown, and the framework any investor, entrepreneur, or operator can steal right now]

4. The Complete Claude /goal Guide for AI Agents 🤖 [What Claude /goal actually is, why most invocations fail, and how to write production-grade goal conditions for financial research, backtesting, and live market monitoring]

5. Circle is building a monopoly in plain sight, but the toll is set by the Fed 🤷‍♂️🏦 [breaking down Circle’s latest Q1 2026 financials, understanding the real economics beneath the headline growth, & figuring out whether Circle deserves a spot in your portfolio + bonus dive into Circle’s Agentic AI bet, which is a clear escape plan]

6. Coinbase’s Q1 2026: the House always wins, until the Casino empties 🤷‍♂️🎰 [deep dive into Coinbase’s Q1 2026, breaking down the key financial facts & figures, what they mean, what’s next & whether it’s worth your time and money now + bonus deep dive into the latest financials of it’s close partner Circle inside]

7. How to Build an AI Agent from Scratch (With Working Code) 🤖 [The design framework, working code, and hard-won lessons that take you from “I want to build an AI Agent” to a working one - in under 60 minutes]

8. Inside Revolut’s PRAGMA: The Foundation Model Trained on 40 Billion Banking Events 🧠 [Architecture, performance benchmarks vs. Stripe, Mastercard, and Visa, regulatory risks, and why PRAGMA may be the most consequential AI bet in consumer finance]

9. Claude Cowork Commands, Scheduled Tasks & Automation Workflows: The Operator’s Playbook 🤖 [Every real slash command, the exact prompts that run deal flow and investor updates on autopilot, and the failure modes that will cost you hours if you don’t know them first]

10. The System for Never Hitting Claude’s Limits 🤖 [Most users burn the majority of their token allocation on architecture mistakes, not actual work. Here’s the operational framework that fixes it]

As for today, here are the 3 fascinating FinTech stories that are changing the world of financial technology as we know it. This was yet another wild week in the financial technology space, so make sure to check all the above stories.

Stripe turned its own workforce into an AI Product Lab 🤖🧪

Following the trends 🧑‍💻 Most companies hand employees a ChatGPT license and call it a transformation. FinTech giant Stripe is now embedding full-time AI practitioners inside its marketing org, one for every 20 people, with a mandate to rewire how each person works from the ground up.

The bet here is definitely not on better tools. It’s on permanently changing human behavior at enterprise scale, and doing it through a model borrowed from one of the most controversial companies in tech. Stripe is now treating its own workforce as a controlled experiment in human-AI collaboration, and the data coming out of that experiment feeds directly into the agentic commerce infrastructure it’s selling to everyone else.

It’s basically a flywheel hiding in a job posting ⚙️

Let’s take a closer look at this, understand what it’s truly all about, why it matters, and why everyone in FinTech, finance, and tech should be paying very close attention.

More on this 👉 The role is called “Forward Deployed AI Accelerator,” borrowed from Palantir’s playbook of embedding engineers inside client organizations to close the gap between software and operations.

Interestingly, Stripe flipped the model inward. Each FDA maps a marketer’s actual workflows, builds custom agents and automations for their specific responsibilities, then coaches them through a structured maturity model: awareness, first win, regular use, full transformation, self-sufficiency.

What’s important to stress here is that success isn’t measured in adoption dashboards. It’s measured in workflows that permanently start with an AI tool. The endgame spelled out in the job spec is super clear: preparing teams to design and oversee autonomous multi-agent systems, not just write better prompts.

Basically, run something similar to Claude Managed Agents inside the org.

Zoom out 🔎 Now, look at the bigger picture and read this alongside Stripe Sessions 2026, where the company shipped 288 products aimed at making AI agents first-class economic actors.

→ Link’s agent wallet covers 250 million users.

→ The Machine Payments Protocol lets agents pay for API calls autonomously.

All while Stripe processed $1.9 trillion in payment volume last year at a whopping $159 billion valuation.

ICYMI: Stripe is building Visa for Machines 🤖💳 [two most important product launches from Stripe Sessions 2026, & how the fintech giant is building the full financial stack for autonomous AI + bonus deep dives into Stripe & how Google just won Agentic Commerce]

Most importantly, the connection between the internal and external plays is the part that most people seem to miss here. Every workflow the FDA team transforms inside Stripe generates real telemetry on how humans and AI agents collaborate in high-stakes enterprise contexts. That operational knowledge thus feeds directly into the agentic infrastructure Stripe sells.

The org itself then becomes a proving ground for the products. Brilliant play 👏

Meanwhile, the protocol wars Stripe was supposed to fight are already over. Seven months after co-building the Agentic Commerce Protocol with OpenAI, Stripe walked across the aisle and joined Google’s Universal Commerce Protocol Tech Council alongside Amazon, Microsoft, Meta, and Salesforce.

THE TAKEAWAY ✈️

What’s next? 🤔 Looking ahead, watch two things from here. First and foremost, whether Google embeds UCP primitives directly into Gemini’s agent infrastructure, giving it a structural advantage over every other AI assistant in commerce. If that happens, Google doesn’t just own the standard. It owns the best implementation of the standard. Second, whether Stripe’s FDA model spreads beyond marketing into risk, compliance, and merchant operations by year-end. Ramp is already running a version for its largest customers. Zooming out, the deeper question now is whether the telemetry from internal transformation creates a durable product advantage, or whether the model quietly dies once the initial cohort reaches self-sufficiency. If agents mature fast enough, the humans who trained alongside them become the most valuable operators in fintech. If they don’t, Stripe just built an expensive coaching program. The early signal from a 70% reduction in content review cycle time suggests the former. But conversion, not adoption, remains the test that matters for everything Stripe is building, inside and out.

ICYMI:

Anthropic is coming for RegTech: Claude won’t sell compliance software - it wants to be the Compliance Layer 😳🤖

The news 🗞️ Dun & Bradstreet just handed Claude the keys to 580 million business entities. And no, that’s not a data partnership. That’s Anthropic positioning itself as the operating system for regulated finance.

The D&B integration is the clearest signal yet that Anthropic is building a vertically integrated compliance infrastructure layer, with KYC/KYB, credit risk, and supplier due diligence running natively inside Claude rather than through the patchwork of middleware vendors that banks and fintechs rely on today.

Let’s break down exactly how the integration works, what it means for RegTech startups competing in the same space, where the competitive pressure lands for OpenAI and Google, and what regulatory response could accelerate or stall the whole thesis.

More on this 👉 Yesterday, D&B announced it would pipe its Commercial Graph directly into Claude through Anthropic’s Model Context Protocol. MCP is the quiet piece that matters here: it’s a standardized connector that gives AI agents auditable, permissioned access to external data while preserving lineage.

Think of it less as an API and more as a governed nervous system for agentic workflows. The integration means Claude can now resolve entities against D&B’s D-U-N-S Number, traverse ownership structures, cross-reference risk signals, and produce regulator-ready documentation. KYC/KYB workflows that took compliance teams days can now be assembled in minutes inside Claude. Let that sink in.

Zoom out 🔎 The timing was surgical here. We must remember that just recently Anthropic dropped ten financial-services agent templates, including a KYC screener that assembles entity files, reviews source documents, and packages escalation reports. D&B showed up as the named data backbone.

Anthropic also announced MCP connectors for Moody’s (credit ratings on 600M+ companies), Verisk (insurance data), and IBISWorld. Add in the $1.5 billion joint venture with Blackstone, Goldman Sachs, and others to embed Claude directly into portfolio companies, and the pattern is clear: Anthropic is building a vertically integrated compliance stack, sold not as software but as infrastructure.

As D&B’s GM of Risk Alex Zuck put it, Claude isn’t just getting more data. It’s getting “the verified context and decision logic required to act.” That distinction between data-as-input and data-as-reasoning is what makes this more than a press release.

Future outlook 🔮 That said, here’s what the RegTech world should be paying attention to.

First and foremost, a large share of compliance middleware exists because KYC/KYB is painful: stitching together vendor portals, normalizing data formats, routing cases through approval chains. If Claude can do that natively, grounded in the same D&B data these platforms already resell, the value proposition of pure-play orchestration vendors erodes fast. The survivors will hence be those with proprietary regulatory logic, jurisdiction-specific interpretation, or deep workflow customization that a general-purpose agent can’t replicate out of the box.

The second-order effect is competitive pressure on other foundation model providers. OpenAI, Google, and xAI now face a gap in regulated verticals. Anthropic’s constitutional AI framework and explainability focus were always a better fit for compliance use cases; locking in D&B and Moody’s as native data partners converts that philosophical advantage into a structural moat.

Matching it therefore requires not just data deals but the governed access layer that MCP provides.

THE TAKEAWAY ✈️

What’s next? 🤔 Looking ahead, watch two things over the next 12 months. First, whether banks actually run production KYC through Claude or treat it as a pilot curiosity. Second, how regulators respond. They’ve been asking for explainability and audit trails in AI-assisted decisions already. Anthropic just gave them exactly that, which could accelerate formal guidance on agentic compliance. If regulators bless the pattern, the adoption curve steepens dramatically. If they balk at autonomous entity verification, even with human-in-the-loop, the whole thesis slows down (maybe even quite dramatically). Either way, the era of selling compliance data in static reports is ending. The money moves to whoever controls the intelligence layer where decisions actually get made. And for now, that layer seems to be ClaudeOS.

ICYMI: Anthropic stopped selling AI to Wall Street and started becoming Wall Street’s Operating System 🏦🧠 [what the 10 finance AI agents actually do, why FactSet dropped 8.1% the same day, what the FIS infrastructure deal and $1.5B Blackstone/Goldman JV mean for distribution & what to expect next + bonus Guide on Claude Managed Agents, Claude Code Routines, & How to Build Your First AI Agent from Sratch inside]

Anthropic’s Claude for Small Business is not just an AI chatbot. It’s a bid to own the whole SMB back office 🤖📊 [what Claude for Small Business really is, why it matters, & what it means for investors, founders, and SaaS companies + bonus dive into how I turned Claude Cowork into my Personal COO, and Anthropic that just told AI founders exactly what to build in 2026]

Circle Agent Stack is an escape plan disguised as a product launch 💸🪙

The news 🗞️ 99% of Circle’s $1.68 billion in 2024 revenue came from just one source: interest on the reserves backing USDC. A single percentage-point rate cut costs them $441 million. Let that sink in 😳

That’s not a business model - it’s basically a weather bet. Yesterday, Circle showed us what the escape plan looks like.

Let’s unpack this.

More on this 👉 On May 11, Circle launched Agent Stack, a five-piece toolkit that lets AI agents hold wallets, discover services, and transact in USDC autonomously.

• Agent Wallets enforce spending rules before transactions execute.

• An Agent Marketplace exposes 500+ service endpoints for programmatic discovery and payment.

• A CLI gives developers deterministic control over wallets, payments, and policies.

• Nanopayments, powered by Circle Gateway, enable gas-free USDC transfers down to $0.000001 by batching thousands of off-chain authorizations into single on-chain settlements.

• And Circle Skills hand AI coding agents the structured context to build on all of it without fumbling USDC’s six-decimal precision.

The timing is very deliberate here. The x402 protocol, the emerging HTTP-native payment standard now governed under the Linux Foundation with Coinbase, AWS, Google, Stripe, Visa, and Anthropic as members, processed $24.24 million in the prior 30 days. USDC settled 99.8% of that volume.

In other words, the market picked USDC as the currency of agent commerce before Circle even shipped the infrastructure. So Circle is now building the plumbing around a choice already made.

But the strategic logic runs deeper than AI hype.

Remember that Circle paid Coinbase and other distribution partners $1.01 billion in 2024. In Q1 2025, Coinbase earned $300 million from its USDC revenue-sharing deal while Circle netted $230 million. Agent Stack is permissionless. Every developer who builds directly on it creates USDC demand that flows through Circle’s own rails, not Coinbase’s.

The agentic economy, where software agents interact programmatically rather than through consumer exchanges, is structurally less dependent on Coinbase’s distribution. That’s not a coincidence - it’s the point, and that’s exactly why Circle is making this bet.

THE TAKEAWAY ✈️

What’s next? 🤔 First and foremost, we must note that the current agentic payment volume is $50 million across 40,000 on-chain agents, which is 0.0001% of stablecoin annual settlement. Circle is spending $570-585 million on platform buildout in 2026 against a market that McKinsey projects at $1 trillion in US agentic commerce by 2030, but that barely exists today. If Stripe’s Machine Payments Protocol captures developers through its $1.9 trillion in existing payment relationships and routes them toward fiat settlement, USDC’s 99.8% share of agent payments could prove temporary. Therefore, when looking ahead, watch two things. First, whether Arc mainnet (Circle’s own L1 with USDC as gas) launches with enough institutional adoption to make the agent-wallet-to-settlement pipeline fully captive. Second, whether the Coinbase revenue-sharing renegotiation in 2026 reflects a Circle that finally has distribution leverage of its own. Most importantly, Agent Stack isn’t a product launch. It’s Circle’s bid to stop being a fund and start being a platform, before the rate cycle makes the decision for them.

ICYMI: Circle is building a monopoly in plain sight, but the toll is set by the Fed 🤷‍♂️🏦 [breaking down Circle’s latest Q1 2026 financials, understanding the real economics beneath the headline growth, & figuring out whether Circle deserves a spot in your portfolio + bonus dive into Circle’s Agentic AI bet, which is a clear escape plan]

🧠 What else I’m watching

• Alipay Enables AI Shopping Delegation 🛒 Alipay has expanded its AI Pay feature, now allowing users to delegate purchases to AI agents. Launched last year, AI Pay has surpassed 100M users and processed over 120M transactions in a single week. Integrated with Alibaba’s Taobao, users can authorize an AI assistant to make purchases on their behalf: after voice commands, Alipay AI Pay generates a one-time delegation for the specific purchase, verifies the user’s identity, and lets the AI monitor prices and complete the order. Alipay plans to extend this capability to recurring scenarios like commuting, utility payments, and repeat orders, further automating everyday transactions. ICYMI:

• Google Integrates BNPL for AI Shopping 🛒 Google is adding Affirm and Klarna Buy Now Pay Later (BNPL) options to Google Pay for users shopping in the Gemini app or Google Search, including AI mode. Shoppers will see Affirm and Klarna buttons at checkout, enabling installment payments. Built on Google’s Universal Commerce Protocol, the integration supports AI-driven shopping experiences. ICYMI: The Android of Commerce - How Google Is Building the Interface Between AI & Money 🤖💸 [why recently introduced WebMCP is a game-changer, how it stacks perfectly into Google’s Android of Commerce playbook, things worth watching, what’s next for FinTechs/Banks/Payments companies, etc. + bonus deep dives into Google’s UCP, how I built an AI OS to run startup with Claude, and how AI is eating software today]

• FinBox’s Atlas Speeds Up Loans with AI 🏦 FinBox has launched Atlas, an AI-native lending platform that slashes loan processing from three weeks to 24 hours. Replacing manual workflows with AI agents, Atlas handles borrower onboarding, document validation, and decision-making. Early results show 85% application completion rates (most in under 10 minutes) and 95%+ first-time-right rates, with turnaround times cut from 21 days to one day and file send-backs reduced by 60%. Atlas Flow uses conversational interfaces across WhatsApp, voice, and video, while Atlas Origin autonomously processes documents, flagging discrepancies in real time. Five financial institutions are already implementing Atlas, with modules for credit appraisal and fraud detection in development. ICYMI:

💸 Following the Money

• Prediction market Kalshi has seen its valuation double to $22B in five months off the back of a $1B Series F funding round.

• Squads raises $18M to build a stablecoin operating system. ICYMI:

• BBVA and Goldman Sachs back new OpenAI company.

👋 That’s it for today! Thank you for reading, and have a relaxing Sunday! And if you enjoyed this newsletter, invite your friends and colleagues to sign up:

Subscribe now

Subscribe now

Every serious AI practitioner working in fintech or tech eventually runs into the same wall. You give an agent a complex task - a sector deep dive, a backtesting pipeline, a daily intelligence digest - and it either halts for confirmation every three steps, loops indefinitely without finishing, or delivers a plausible-looking output that quietly fails to meet the actual requirements.

The bottleneck is rarely the model. It’s the spec.

/goal is Claude Code’s mechanism for turning a session into an autonomous loop: the agent runs, verifies whether the goal condition is met, and continues until it is - without checking in with you.

When used correctly, you define the end state, configure the harness, and return to finished work. When used naively - which is how most people use it - you return to either a frozen session or a confident-sounding mess.

This guide therefore covers:

• How /goal actually evaluates conditions (and why most conditions are unevaluable)

• The three-element formula for writing conditions that work

• The reliability architecture that determines whether a multi-hour agent run completes, and

• Three production-grade prompt templates built for the workflows that matter most in fintech - deep competitive research, code-heavy builds like backtests and dashboards, and ongoing portfolio and market monitoring.

• Bonus: Two companion pieces that pair directly with the templates above - The System for Never Hitting Claude’s Limits, covering the context engineering and model selection discipline that keeps long /goal runs from burning your allocation, and an End-to-End Guide to Claude Code Routines, with copy-paste routines for autonomous code review, deploy verification, and deal flow screening.

It also addresses a widespread misconception that “zero hand-holding” is an engineering outcome rather than a marketing claim. Context rot silently degrades long runs well before the window fills. Anthropic’s own research on long-running agents shows that reliability comes from the harness, not the model.

And for fintech founders & operators, what every generic framework omits - data sensitivity tiers, environment segregation, regulatory output flagging, simulation-only constraints - can mean the difference between a useful research agent and an operational liability.

Let’s dive in.

What /goal Actually Is (And What It Isn’t)

Read more

Good morning & happy Thursday! Today, we’re diving deep into Anthropic, which now wants to become the operating layer between every SaaS tool an SME touches (what Claude for Small Business really is, why it matters, & what it means for investors, founders, and SaaS companies + bonus dive into how I turned Claude Cowork into my Personal COO, and Anthropic that just told AI founders exactly what to build in 2026), and Coinbase, which is quietly building a financial superplatform in a difficult market (deep dive into Coinbase’s Q1 2026, breaking down the key financial facts & figures, what they mean, what’s next & whether it’s worth your time and money now + bonus deep dive into the latest financials of it’s close partner Circle inside). So let’s jump straight into the fascinating stuff 🌶️

Anthropic’s Claude for Small Business is not just an AI chatbot. It’s a bid to own the whole SMB back office 🤖📊

The BIG News 🔥 Anthropic simply cannot stop shipping… The AI giant just made its clearest move yet to become the operating layer between every SaaS tool a small business touches, and the playbook looks less like an AI company than like a payroll provider.

Claude for Small Business, launched May 13 inside Claude Cowork, ships 15 agentic workflows and 15 repeatable skills that plug into QuickBooks, PayPal, HubSpot, Canva, DocuSign, Google Workspace, and Microsoft 365.

No new software to learn. No engineers required. You just toggle it on, pick a job, and Claude executes multi-step work across your entire stack - then pauses for human approval before anything gets sent, paid, or posted.

The target market is 36 million U.S. small businesses generating 44% of GDP, nearly all of which lack the headcount or budget for the kind of orchestrated automation that Fortune 500 companies pay seven figures for. If this works, Anthropic doesn’t just sell AI to small businesses. ClaudeOS becomes the connective tissue between every app they already run 🤖

Let’s dive deeper into this and break down exactly what that means for investors, founders, and the SaaS companies caught in the middle.

Read more

Good morning & happy Wednesday! Today’s issue is incredibly packed as we’re diving into Anthropic, which is now coming for RegTech (what Dun & Bradstreet integration in Claude is really all about, what it means for RegTech startups competing in the same space, & what’s next + bonus deep dives into how Claude is becoming Wall Street’s Operating System & Anthropic that just told AI founders exactly what to build in 2026), and Circle that’s building a monopoly in plain sight, yet the toll is set by the Fed (breaking down Circle’s latest Q1 2026 financials, understanding the real economics beneath the headline growth, & figuring out whether Circle deserves a spot in your portfolio + bonus dive into Circle’s Agentic AI bet, which is a clear escape plan). So let’s jump straight into the finnovative stuff 🌶️

Anthropic is coming for RegTech: Claude won’t sell compliance software - it wants to be the Compliance Layer 😳🤖

The news 🗞️ Dun & Bradstreet just handed Claude the keys to 580 million business entities. And no, that’s not a data partnership. That’s Anthropic positioning itself as the operating system for regulated finance.

The D&B integration is the clearest signal yet that Anthropic is building a vertically integrated compliance infrastructure layer, with KYC/KYB, credit risk, and supplier due diligence running natively inside Claude rather than through the patchwork of middleware vendors that banks and fintechs rely on today.

Let’s break down exactly how the integration works, what it means for RegTech startups competing in the same space, where the competitive pressure lands for OpenAI and Google, and what regulatory response could accelerate or stall the whole thesis.

Read more

Good morning & happy Tuesday! Today, we’re diving into Stripe, which is now turning its workforce into an AI Product Lab (what their Forward Deployed AI Accelerators are all about, and why everyone should be paying attention + bonus deep dive into how Stripe is building Visa for Machines, how to deploy multi-agent systems in enterprises at scale & Anthropic that just told AI founders exactly what to build in 2026 inside), and Circle’s Agent Stack, which is actually an escape plan disguised as a product launch (what Circle is really trying to do here and why it makes a ton of sense + bonus dive into Google that just won Agentic Commerce & how to build an AI Agent from scratch inside). So let’s jump straight into the interesting stuff 🌶️

Stripe is turning its own workforce into an AI Product Lab 🤖🧪

Following the trends 🧑‍💻 Most companies hand employees a ChatGPT license and call it a transformation. FinTech giant Stripe is now embedding full-time AI practitioners inside its marketing org, one for every 20 people, with a mandate to rewire how each person works from the ground up.

The bet here is definitely not on better tools. It’s on permanently changing human behavior at enterprise scale, and doing it through a model borrowed from one of the most controversial companies in tech. Stripe is now treating its own workforce as a controlled experiment in human-AI collaboration, and the data coming out of that experiment feeds directly into the agentic commerce infrastructure it’s selling to everyone else.

It’s basically a flywheel hiding in a job posting ⚙️

Let’s take a closer look at this, understand what it’s truly all about, why it matters, and why everyone in FinTech, finance, and tech should be paying very close attention.

Read more

Subscribe now

Most people using Claude Opus 4.7 are leaving the majority of the model’s capability on the table.

Not because the model is limited. Claude Opus 4.7 - released April 16, 2026 - is Anthropic’s most capable generally available model. It reasons more precisely, follows instructions more literally, and executes long-horizon autonomous tasks more reliably than any previous Claude release.

The problem is prompting.

People copy prompts written for ChatGPT. They carry over habits from Opus 4.6 that no longer apply. They add vague instructions like “be concise” or “think carefully” without understanding what those phrases actually trigger in the new architecture. And they skip the most powerful lever the model offers: the effort parameter, which didn’t exist before Opus 4.7 in its current form, and now controls intelligence allocation more than any other variable.

Opus 4.7 is a different animal from its predecessors. It is more literal - it will not silently generalize an instruction you wrote for one item across all items. It respects effort levels strictly, meaning a low-effort prompt genuinely produces a low-effort response. Its default tone is more direct and opinionated, with fewer of the warm, hedging qualifiers that characterized Opus 4.6. Its design defaults, its code review behavior, its tool-use patterns - all have shifted in ways that reward prompts specifically calibrated to how this model thinks.

This guide is built for founders who use Claude to reason through hard problems, operators who run processes on it, and investors who analyze markets with it. It distills Anthropic’s official documentation (31-page prompting PDF) into a practical framework, then puts that framework to work in 10 Mega Prompts engineered for Claude Opus 4.7, covering the highest-leverage tasks for anyone building or backing a business.

Below is everything you need to get the best output this model can produce, every time.

Step 0: Choose the Right Model

Before writing a single word of your prompt, choose the right model. Anthropic’s current lineup includes Claude Opus 4.7, Claude Sonnet 4.6, and Claude Haiku 4.5, each optimized for different workloads.

Claude Opus 4.7 is the flagship. It excels at long-horizon agentic work, complex reasoning, knowledge work, high-resolution vision, and memory-intensive tasks. It costs more and runs slower via the API ($5/MTok input, $25/MTok output), but when the task demands real thinking - strategic analysis, multi-step research, complex code, autonomous execution - nothing else in the generally available lineup comes close.

Claude Sonnet 4.6 is the balanced workhorse: strong reasoning at a faster speed and lower cost. For most everyday tasks, Sonnet covers roughly 80% of use cases well.

Claude Haiku 4.5 is the speed specialist: fastest, cheapest, ideal for high-volume, straightforward tasks like classification, extraction, and summarization.

💡 The rule of thumb: start with Sonnet. Move to Opus when you need genuine cognitive depth. Drop to Haiku when speed matters more than intelligence.

The Framework: Prompting Claude Opus 4.7

Anthropic’s documentation describes a hierarchy of techniques ordered by impact. Most people jump straight to advanced techniques and skip the fundamentals. That’s backwards. Here’s the framework in the order that matters, updated for the specific behaviors of Opus 4.7.

1. Set the Effort Level First

This is new in its current form to Opus 4.7, and it’s the most important thing in this guide.

The effort parameter controls how much intelligence the model applies to a task. Unlike earlier versions, where the model would often overperform regardless of instructions, Opus 4.7 respects effort levels strictly - especially at the low end. As Hex’s CTO put it: “low-effort Opus 4.7 is roughly equivalent to medium-effort Opus 4.6.” Set effort deliberately for every use case.

In the API:

client.messages.create(
    model="claude-opus-4-7",
    max_tokens=64000,
    thinking={"type": "adaptive"},
    output_config={"effort": "xhigh"},
    messages=[{"role": "user", "content": "..."}],
)

When running at max or xhigh, set your max token budget to at least 64k - the model needs room to think, reason, and execute across tool calls and subagents.

Practical rule: If you see shallow reasoning on a complex task, raise the effort level before changing your prompt. That’s almost always the correct fix..

2. Be Specific and Explicit - More Than You Think

This sounds obvious. In Opus 4.7, it’s critical in a way it wasn’t before.

The single highest-leverage prompting move is specificity. Claude Opus 4.7 is extremely capable but also extremely literal. Vague prompts get scoped, not generalized. If you want an instruction applied across all sections, not just the first, you have to say so. If you want “above and beyond” output, you have to request it explicitly.

Anthropic’s own recommendation: think of your prompt as instructions to a brilliant but literal new hire on their first day. They’ll do exactly what you say - so say exactly what you mean.

❌ Weak prompt:

Write about market positioning

✅ Strong prompt:

Analyze the 3 most effective market positioning strategies for B2B SaaS companies targeting mid-market in a crowded category. For each strategy, explain what's driving its effectiveness, provide one specific company example, and assess whether it's likely to strengthen or weaken over the next 18 months. Apply this framework to all three strategies, not just the first.

The difference isn’t more words. It’s more specificity, explicit scope, and a format directive at the end.

3. Structure Prompts with XML Tags

This is Claude’s structural superpower, and almost nobody uses it correctly.

Claude was specifically trained to recognize XML tags as structural markers. When your prompt has multiple components - context, instructions, data, constraints, output format - XML tags prevent Claude from mixing them up. Here’s what that looks like for a strategic decision:

<context>
You are helping me evaluate a potential Series A investment.
The company is a vertical SaaS business targeting logistics operators, currently at $2.4M ARR growing 15% MoM.
</context>

<instructions>
Analyze the three key risks that most commonly derail vertical SaaS companies at this stage.
For each risk, explain the warning signs and what a founder should be doing to mitigate them.
Apply this analysis to all three risks, not just the most obvious one.
</instructions>

<constraints>
- Be direct. Give me your honest assessment, not a balanced "it depends."
- Use specific examples from real companies where possible.
- Flag any assumptions you're making.
- Maximum 600 words.
</constraints>

Claude sees the tags and immediately understands that <context> is background (not a task to respond to), <instructions> is the actual task, and <constraints> are the guardrails.

Tag names are flexible - there’s no magic set of “correct” tags. Use whatever makes semantic sense: <background>, <rules>, <examples>, <output_format>. Consistency across your prompts matters more than the specific names you choose.

4. Show Claude What Good Looks Like

If one technique consistently separates good outputs from great ones, it’s this: show Claude what good looks like.

Instead of describing the tone, format, or style you want in abstract terms, provide two to three concrete examples. Claude will pattern-match against these far more reliably than it will follow descriptive instructions alone. Wrap examples in <example> tags (multiple examples in <examples> tags) so Claude distinguishes them from instructions.

<examples>
  <example>
    Input: "We need to cut 20% of the engineering budget"
    Output: "Reducing engineering spend by 20% requires prioritization across three areas: contractor headcount, infrastructure costs, and tooling licenses. Here's a phased approach that preserves our two highest-impact product initiatives..."
  </example>
</examples>

Now analyze this situation using the same approach:
"We need to extend our runway by 6 months without reducing headcount"

Anthropic recommends 3–5 examples for best results. You can also ask Claude to evaluate your examples for relevance and diversity, or to generate additional ones based on your initial set.

5. Build in Reasoning Steps

For complex problems requiring analysis, multi-step reasoning, or strategic judgment, telling Claude to work through its reasoning before producing a final answer dramatically improves accuracy.

The simplest version: add “Think through this step by step before giving your final answer” to your prompt.

The more structured version uses tags to separate reasoning from output:

<instructions>
Evaluate whether we should expand into the Southeast Asian market this year.

Before giving your recommendation, work through the analysis inside <analysis> tags. Consider: market size and growth trajectory, regulatory requirements by country, competitive landscape, our current operational capacity, and capital requirements vs. expected payback period.

Then provide your final recommendation with a clear resource allocation suggestion.
</instructions>

Forcing visible reasoning prevents Claude from pattern-matching to the most likely answer and back-filling justification after the fact.

Adaptive thinking in Opus 4.7: Claude now dynamically decides when and how much to think based on the effort setting and task complexity. At high and xhigh effort, deep reasoning is largely automatic for demanding tasks. Extended thinking with a fixed budget_tokens is no longer supported - adaptive thinking is the only thinking-on mode, and Anthropic’s internal evaluations show it reliably outperforms the old fixed-budget approach. To guide reasoning explicitly:

After receiving results, carefully reflect on their quality and determine optimal next steps before proceeding. Use your thinking to plan and iterate based on this new information, then take the best next action.

6. Load Rich Context

Claude can only work with what you give it. The more relevant context you include, the more tailored and accurate the output becomes.

Upload documents. Paste data. Provide company background. Share your goals. Explain your audience. Don’t make Claude guess what you already know.

<background>
Our company builds financial infrastructure for neobanks in emerging markets.
We're Series B, $12M ARR, primarily serving West Africa and Southeast Asia.
Our main competitors are Banking-as-a-Service players like Railsbank, Synapse, and local incumbents.
We differentiate on compliance coverage and local payment rail integrations.
</background>

<data>
[Paste your Q1 metrics, customer feedback, churn data, or whatever's relevant]
</data>

<task>
Based on this context, identify our three biggest growth opportunities for the next two quarters.
</task>

For long documents (20k+ tokens), put the documents at the top of your prompt, above your instructions and query. Queries at the end can improve response quality by up to 30% in tests with complex, multi-document inputs.

7. Specify the Output Format

Don’t leave the structure of Claude’s response to chance. If you want a table, ask for a table. If you want a specific word count, state it. If you want an executive brief with defined sections, describe each section.

<output_format>
Respond with:
1. A one-paragraph executive summary (3-4 sentences max)
2. A comparison table with columns: Factor | Current State | Target State | Gap
3. A "Recommended Actions" section with 3 specific next steps, ranked by impact
</output_format>

Explicit format specs eliminate the most common frustration people have with AI: getting a 2,000-word essay when you wanted a concise brief, or getting bullet points when you needed flowing analysis.

8. Define Constraints - Especially “Don’ts”

Telling Claude what not to do is just as important as telling it what to do. Without constraints, Claude defaults to its training patterns, which often means overly formal, hedge-heavy output that sounds like a committee wrote it.

<constraints>
- Do NOT open with "In today's rapidly evolving landscape" or any variant
- Skip the preamble. Start with the most important insight.
- No bullet points — write in prose paragraphs
- If you're uncertain about a claim, flag it explicitly rather than hedging everything
- Maximum 500 words
- Be direct. I want your honest assessment, not a balanced "it depends."
</constraints>

One note specific to Opus 4.7: the model is more direct and opinionated than 4.6, with fewer of the warm, validating preambles (”Great question!”). If you want a warmer tone, request it explicitly. If you want the raw, direct output, the model defaults to - you’re already there.

9. Control Verbosity

Opus 4.7 calibrates response length to how complex it judges the task to be - shorter on simple lookups, much longer on open-ended analysis. If your use case requires a specific verbosity, tune it explicitly.

To decrease verbosity:

Provide concise, focused responses. Skip non-essential context and keep examples minimal.

Key insight from Anthropic: Positive examples showing appropriate concision work better than negative instructions like “don’t be verbose.” Show Claude a response at the length and depth you want, and it will match that pattern far more reliably than it will follow abstract length instructions.

10. Putting It All Together

Here’s what a properly structured Claude Opus 4.7 prompt looks like when you combine all nine techniques:

<context>
I'm the CEO of a B2B fintech startup ($8M ARR, 45 employees).
We're deciding whether to raise a Series B now or extend runway and raise in 18 months.
Current runway: 14 months. Revenue growth: 12% MoM. CAC payback: 8 months.
</context>

<instructions>
Analyze both timing options. Before giving your recommendation, work through the
trade-offs in <analysis> tags, considering:
- Current market conditions for fintech Series B rounds
- Our specific metrics relative to typical Series B benchmarks
- The risk/reward of raising now vs. in 18 months at potentially better metrics
- What we should use the 18 months to optimize if we extend

Then provide a clear recommendation with a specific action plan.
Apply your analysis to both options equally — don't weight one by default.
</instructions>

<constraints>
- Be direct. Give me your honest read, not a balanced "it depends."
- Use specific benchmarks from comparable fintech Series B raises where possible.
- Flag any assumptions you're making about market conditions.
- Keep the total response under 700 words.
</constraints>

<output_format>
1. Analysis (in <analysis> tags)
2. Recommendation (2-3 sentences, clear and direct)
3. 90-day action plan (5 specific actions, whether we raise now or extend)
</output_format>

This prompt is clear, structured, specific, and constrained. It tells Claude exactly what to do, how to think about it, what to avoid, and how to format the response. The output from this kind of prompt is categorically different from “Should I raise a Series B now or wait?”

The 10 Claude Opus 4.7 Mega Prompts

Understanding the principles is one thing. Having them already engineered into prompts you can use today is another.

Below are 10 Mega Prompts, each structured using every technique above and calibrated for Claude Opus 4.7’s specific behaviors - its literalism, effort sensitivity, and adaptive thinking. Each covers a high-leverage task for founders, operators, and investors.

How to use these prompts:

• Fill in every [PLACEHOLDER] before submitting. Blank placeholders produce generic output.

• Enable web search for prompts that reference real-time data: Research (#1), Competitor Analysis (#8), Stock/Deal Analysis (#9), and Reverse-Engineering (#10). Without it, Claude will rely on training data and flag confidence gaps.

• Use xhigh or high effort for all 10 prompts. For #9 and #10, start at xhigh or max.

• Mix and match: you can add chain-of-thought <analysis> tags, output constraints, or anti-patterns from one prompt into any other.

• Don’t want to build prompts from scratch? I’ve created a custom Claude skill that converts any prompt into an Opus 4.7-optimized version automatically. Grab it at the end of this guide.

#1 - Deep Research

Produces an executive brief with sourced findings, expert perspectives, and contrarian views.

Read more

If you’re not a subscriber, here’s what you missed this week:

1. Anthropic Just Told AI Founders Exactly What to Build in 2026 🦄 [1 million conversations. 9 consumer AI domains. A full founder playbook - plus where Anthropic’s own products will and won’t compete]

2. He Got Fired From OpenAI at 22. Then He Turned $225M Into $5.5B by Betting on the AI Infrastructure No One Else Was Watching 💸 [Inside Leopold Aschenbrenner’s Situational Awareness Fund: the AGI investment thesis, the full portfolio breakdown, and the framework any investor, entrepreneur, or operator can steal right now]

3. Turn ChatGPT Images 2.0 & Claude Design Into Your Chief Designer 🫟 [How to build a Brand Identity, Design System, and Product Prototype in hours - without a designer. It’s the best 2026 workflow for FinTech & AI founders who need institutional-grade design fast]

4. Anthropic stopped selling AI to Wall Street and started becoming Wall Street’s Operating System 🏦🧠 [what the 10 finance AI agents actually do, why FactSet dropped 8.1% the same day, what the FIS infrastructure deal and $1.5B Blackstone/Goldman JV mean for distribution & what to expect next + bonus Guide on Claude Managed Agents, Claude Code Routines, & How to Build Your First AI Agent from Sratch inside]

5. A 3-year-old firm nobody’s heard of just bought the world’s largest corporate travel company 😳✈️ [why Long Lake Management is acquiring Amex GBT, and what it means for Ramp, Brex and Navan + bonus deep dives into American Express and why it skipped the Agent Protocol wars & Anthropic, which just told founders exactly what AI products to build in 2026]

6. Citi’s AI Avatar vs. the $650 billion FinTech wave it can’t outrun 🤖🌊 [AI Avatar Citi Sky, Arc Platform, how it stacks against AI initiatives from other major banks, and why it might be impossible to outrun FinTech competitors + bonus deep dive into Revolut’s AI Model for Money & how Anthropic is building Wall Street’s AI operating system inside]

7. How to Build an AI Agent from Scratch (With Working Code) 🤖 [The design framework, working code, and hard-won lessons that take you from “I want to build an AI Agent” to a working one - in under 60 minutes]

8. Inside Revolut’s PRAGMA: The Foundation Model Trained on 40 Billion Banking Events 🧠 [Architecture, performance benchmarks vs. Stripe, Mastercard, and Visa, regulatory risks, and why PRAGMA may be the most consequential AI bet in consumer finance]

9. Claude Cowork Commands, Scheduled Tasks & Automation Workflows: The Operator’s Playbook 🤖 [Every real slash command, the exact prompts that run deal flow and investor updates on autopilot, and the failure modes that will cost you hours if you don’t know them first]

10. The System for Never Hitting Claude’s Limits 🤖 [Most users burn the majority of their token allocation on architecture mistakes, not actual work. Here’s the operational framework that fixes it]

As for today, here are the 3 incredible FinTech stories that are transforming the world of financial technology as we know it. This was yet another insane week in the financial technology space, so make sure to check all the above stories.

OpenAI & Anthropic outsourced their sales teams to private equity 🤑💸

The news 🗞️ The two leading AI labs stopped pretending software sells itself.

On the same day, OpenAI and Anthropic each finalized & signed joint ventures with the biggest names in private equity - TPG, Blackstone, Goldman Sachs, and Bain Capital - to physically embed their engineers inside thousands of portfolio companies.

The combined capital committed: over $5.5 billion. The model: Palantir’s forward-deployed playbook, scaled to the entire mid-market economy in one move.

One of these deals includes a 17.5% guaranteed annual return to investors, which is either a masterstroke or a ticking liability on top of $14 billion in projected losses. The other offers no guarantee at all, and that contrast tells you almost everything about where these two companies stand heading into their IPOs. For fintech operators and financial services investors, Goldman’s involvement on the Anthropic side opens a direct pipeline into wealth management, lending, and insurance portfolios that most startups spend years trying to reach. And that could be huge.

Let’s break down the deal structures, the real competitive implications for enterprise software & consulting, and the 3 specific things to watch over the next 12 months.

More on this 👉 Within hours of each other on May 4, OpenAI and Anthropic each finalized joint ventures with private equity giants to embed their models directly inside portfolio companies.

OpenAI’s vehicle, called The Deployment Company, raised over $4 billion from a 19-firm consortium led by TPG, valued at $10 billion pre-money. Anthropic countered with a $1.5 billion venture anchored by Blackstone, Hellman & Friedman, and Goldman Sachs.

Both are borrowing Palantir’s playbook: forward-deployed engineers sitting inside client organizations, wiring AI into actual workflows rather than shipping licenses and hoping for the best.

The structural differences matter here.

→ OpenAI offered PE investors a 17.5% guaranteed annual return floor over five years. At $4 billion in committed capital, that’s up to $700 million a year in guaranteed payouts if the venture underperforms, stacked on top of projected $14 billion in 2026 losses.

→ Anthropic offered ordinary equity with no floor.

Read that as a confidence gap, a cultural signal, or both: the safety-focused lab apparently thinks the commercial upside speaks for itself.

Zoom out 🔎 For fintech and financial services, the Goldman Sachs piece is the one to watch. Goldman is investing through its asset management unit, immediately connecting Anthropic’s engineers to a portfolio spanning wealth management, lending, and insurance. Blackstone, with $1.3 trillion in assets, adds thousands more companies across every sector where financial products move.

If those portfolio companies get AI-native operations 18 months before their competitors, the gap won’t close easily. For a company like Revolut, which built its edge on engineering speed and lean operations, the question flips: what happens when legacy banks and mid-market financial services firms suddenly get Anthropic engineers rebuilding their back offices?

ICYMI: Inside Revolut’s PRAGMA: The Foundation Model Trained on 40 Billion Banking Events 🧠 [Architecture, performance benchmarks vs. Stripe, Mastercard, and Visa, regulatory risks, and why PRAGMA may be the most consequential AI bet in consumer finance]

The second-order effect most people will miss is what these deals do to enterprise software incumbents. RBC flagged earlier this year that Anthropic product launches sent Salesforce, Workday, and Intuit shares down 6-13%.

Now imagine that pressure applied not through an open market but through PE board mandates. When Blackstone tells a portfolio company to deploy Claude across its operations, the CRM vendor doesn’t get a renewal conversation. It gets a replacement conversation.

ICYMI:

THE TAKEAWAY ✈️

What’s next? 🤔 Looking ahead, there are three key things to track over the next twelve months. First, whether OpenAI’s 17.5% guarantee becomes a template or a cautionary tale. Second, whether the forward-deployed engineer model can actually scale - Palantir took two decades to prove it, and talent that can code and navigate corporate politics remains scarce. Third, watch Google. DeepMind partnered with McKinsey, BCG, and Deloitte in April, betting on consultants as distribution rather than competing with them. One of these two approaches will definitely win. We just don’t know which one yet.

ICYMI: OpenAI doesn’t want to build a Phone. It wants to build the last App Store 😳📱 [why it’s not about killing the iPhone here but all about building the last App Store + bonus deep dive into OpenAI’s OS strategy & the quest to become the ultimate AI Super App]

OpenAI is building a personal CFO in plain sight 🤖📊 [what their Hiro Finance M&A is all about, why they needed them & how it stacks into their bigger long-term strategy + bonus deep dive into OpenAI’s Agentic Singularity strategy inside]

One unreleased AI model just triggered a global financial emergency 😳🚨 [why Claude Mythos prompted the US Treasury, the Fed, the Bank of Canada, and UK financial regulators to hold emergency meetings with their largest banks, where the biggest risks are & what’s next + bonus deep dive into Anthropic’s leaked Claude source code & how I turned Claude Code into my 10X engineer]

Anthropic stopped selling Intelligence and started selling Infrastructure 🤖💸 [breaking down Claude Managed Agents, why the architecture is really a platform lock-in strategy disguised as a product launch, which startups just lost their reason to exist + how to solve AI Agent context window using Skill Graphs & deep dive into Anthropic’s leaked Claude Source inside]

Coinbase’s 3rd layoff in 4 years isn’t an AI Story. It’s a Business Model Story 🤖🙅‍♂️

The news 🗞️ Coinbase has now cut 18% of its workforce in 2022, 20% in 2023, and 14% yesterday. Each round followed the same script: hire aggressively during a bull market, get caught by a downturn, slash headcount with the vocabulary of the moment 🤷‍♂️

In 2022 and 2023, it was “we over-hired.” In 2026, it’s “AI is reshaping how we work.” The pattern is identical. Only the framing has changed.

Let’s unpack this.

More on this 👉 The 700-person cut, announced May 5, followed a Q4 2025 in which transaction revenue fell 37% year-over-year and a Q1 2026 in which adjusted EBITDA collapsed 51%. Full-year 2025 operating expenses grew 35% while revenue grew 9%. Headcount rose 31%. In a single year 👀

That math clearly doesn’t math with or without AI. Brian Armstrong’s memo acknowledged both crypto weakness and AI productivity as drivers, but his own language gives the game away: he described the restructuring as preparation for “our next phase of growth,” which is cyclical thinking dressed in structural clothing.

If this were truly an AI-native redesign, you wouldn’t frame it as a bridge to the next bull run.

Zoom out 🔎 We must note that Armstrong isn’t alone in reaching for the AI explanation.

→ Block cut 40% of its workforce in February, citing AI.

→ Snap cut 16% in April.

→ Atlassian cut 10% in March.

Sam Altman himself called it out at the India AI Impact Summit: companies are blaming AI for layoffs they’d be making anyway. A Harvard Business Review study found that only 2% of C-suite executives said their cuts were actually driven by existing AI capabilities. Gartner predicts that half the companies attributing headcount reductions to AI will rehire for similar roles by 2027.

That said, the productivity gains definitely aren’t fiction. Block’s internal AI platform “Goose” drove 40% more code per engineer in six months, and the company released more Square products in Q4 2025 than in all of 2024. Coinbase has mandated GitHub Copilot and Cursor adoption with a target of 50% AI-written code. Engineers are genuinely shipping faster. But the distance between “developers are more productive with AI tools” and “we need to eliminate 4,000 jobs” is vast, and the gap is filled mostly by overhiring corrections and market weakness.

For Coinbase specifically, the AI story obscures a more pressing competitive problem. The company ranked eighth globally among crypto exchanges with 6.1% market share as of early 2026, far behind Binance at 39.2%. Kraken, Gemini, and Bullish have all gone or are going public, eroding Coinbase’s position as the only major listed US exchange. That said, the fee compression is coming regardless of how flat the org chart gets 🤷‍♂️

THE TAKEAWAY ✈️

What’s next? 🤔 Looking ahead, I’d watch two things over the next 12 months. First and foremost, whether Coinbase starts quietly rehiring in late 2026 or early 2027 if crypto recovers, which would confirm this was cyclical cost-cutting, not structural transformation. Second, whether Armstrong’s “one-person team” experiment survives contact with the compliance, risk, and customer support demands of a regulated financial services company. Remember that Klarna tried automating 700 support roles with AI in 2024, then reversed course a year later when the CEO admitted they’d moved too aggressively. Because regulated fintech is exactly where the “AI replaces headcount” thesis gets its real stress test.

ICYMI:

Crypto’s smartest money is now betting AI Agents need blockchain banks 🤖🏦

Following the money 💸 The two crypto VCs with the best infrastructure track records just told you where the money’s going next, and it’s not where the “crypto is dead” crowd thinks.

Let’s take a look at this.

More on this 👉 Paradigm is raising $1.5 billion for a fund that formally adds AI, robotics, and frontier tech to its mandate. Haun Ventures closed $1 billion across two new funds on May 4, with AI agents sitting alongside stablecoins and tokenization as core theses.

Between them, that’s $2.5 billion in fresh capital pointed at a single bet: autonomous AI agents will need financial rails that legacy payments can’t provide, and crypto infrastructure is the answer.

Zoom out 🔎 The easy read is that crypto VCs are chasing AI’s heat. The better read is that the chase already happened at the portfolio level.

Silicon Valley Bank data shows 40 cents of every crypto VC dollar in 2025 went to companies blending AI and crypto, up from 18 cents a year earlier. Paradigm and Haun are formalizing what their deal flow already told them. Haun’s prior fund proved the thesis works in practice: Bridge sold to Stripe for $1.1 billion and BVNK to Mastercard for $1.8 billion, both stablecoin payment infrastructure plays that look prescient now that agent-to-agent payment protocols from Coinbase, Google, and Stripe are live or nearly so.

We must note that the firms are attacking different layers. Paradigm is going deep on protocol infrastructure, backing decentralized AI training (Nous Research, $50M Series A) and co-building smart contract security tools with OpenAI. Haun is focused on the regulated financial plumbing: her largest position is Erebor, Palmer Luckey’s FDIC-insured digital bank valued at $4.35 billion.

One builds the pipes, the other builds the on-ramps.

Not everyone agrees convergence is the play, though. a16z crypto raised $2.2 billion the same week, dedicated 100% to crypto. That divergence matters for founders shopping term sheets: pitch the agent economy to Paradigm and Haun, pitch pure protocol infrastructure to a16z.

THE TAKEAWAY ✈️

What’s next? 🤔 Looking ahead, the real question now is not whether AI agents will transact autonomously. It’s whether they’ll do it on crypto rails or through patched-together traditional payment APIs. Machine-driven transactions are projected to touch $9 trillion annually by 2030. If even a fraction settles on-chain, the infrastructure companies being funded right now become the Visa and SWIFT of the agent economy. Thus, for founders: design for machine customers from day one, build compliance into the architecture, and know that the capital is there if you’re building at the overlap. Meanwhile, for investors watching from the sidelines, the window to take a position on this convergence is narrowing fast.

ICYMI:

🧠 What else I’m watching

• Lloyds Deploys AI Agent Platform 🤖 Lloyds Banking Group has launched Envoy, a Google Cloud-powered platform enabling teams to train, use, and share AI agents securely across its operations. The platform offers ready-to-use templates for building task-specific agents, which can be reused and shared via an internal marketplace. Envoy is said to have the potential to boost productivity, enhance customer journeys, and drive innovative business models. Agents can retain customer data in memory to support ongoing interactions, with built-in compliance, safety checks, and human oversight ensuring responsible use. ICYMI:

• NatWest Brings Mortgages to ChatGPT 🏡 NatWest has launched an app on ChatGPT, allowing users - both customers and non-customers - to tag the bank in queries and receive tailored mortgage and home-buying guidance without leaving the platform. Powered by NatWest APIs, users can calculate borrowing capacity, test affordability and deposit scenarios, and access personalized mortgage rates. They are then directed to NatWest’s channels to book appointments or start digital mortgage applications. This follows NatWest’s 2025 strategic partnership with OpenAI, granting early access to OpenAI’s tools and bespoke services. ICYMI:

• Visa Expands Agentic Ready Program 🌍 Visa is rolling out its Agentic Ready program in Canada and Malaysia, helping payments ecosystems prepare for AI agents initiating and completing transactions on behalf of consumers. In Canada, all Big 5 banks - BMO, CIBC, RBC, Scotiabank, and TD - are onboard as issuing partners, with more expected to join. The program allows participants to test agent-initiated payments in real-world environments using live cards and merchants, validating payment flows and trust mechanisms. In Malaysia, Alliance Bank, CIMB, and Maybank are already participating. ICYMI:

💸 Following the Money

• Fun, a payments infrastructure company used by the likes of Polymarket for crypto onramping, has emerged from stealth with $72M in Series A funding.

• MoonPay, the leading crypto payments network, announced the acquisition of DFlow, the fastest-growing trading infrastructure platform on Solana.

• Core Scientific to acquire bitcoin miner Polaris in $421M deal to expand Oklahoma AI data center campus

👋 That’s it for today! Thank you for reading, and have a relaxing Sunday! And if you enjoyed this newsletter, invite your friends and colleagues to sign up:

Subscribe now

Subscribe now

You are paying for Claude. You are hitting limits anyway.

Maybe it happens mid-project, 45 minutes into a complex research session, right when the work is getting good. Maybe it is the second time in a single morning. Maybe you upgraded from Pro to Max and still hit the ceiling within a day. The experience is always the same: the wall appears, the session dies, and the context you built up is gone.

You are not alone.

In early 2026, Anthropic acknowledged that users were hitting usage limits in Claude Code way faster than expected and called fixing it the “top priority for the team.” A GitHub bug report from March 31, 2026 drew hundreds of comments from Pro and Max subscribers, including Max 20x users at $200/month, watching single prompts consume 10–20% of their daily allocation.

Anthropic has since expanded compute capacity, including a partnership with SpaceX announced May 6, 2026, which doubled Claude Code’s five-hour rate limits and removed peak-hour throttling for Pro and Max accounts. Limits are genuinely better now than they were in Q1 2026.

But even with doubled limits, most users are burning the majority of their allocation on architecture mistakes, not actual work. Long conversations that re-tokenize thousands of words of history on every message. Broad file reads when only one function mattered. Defaulting to Opus when Sonnet would have handled it identically. Prompting Claude to “fix the whole thing” when only section 3 was wrong.

The operators who never hit limits are not on bigger plans. They have built a system: persistent knowledge, decomposed workflows, and a model selection discipline that delivers the same output at a fraction of the context cost.

This guide is that system.

Part 1: How Claude’s Limits Actually Work

Before you can beat the system, you need to understand what you are actually fighting.

The Three-Layer Limit Stack

Read more