In-Depth · Xiaohu Explains

Anthropic AI-Native Founder's Playbook: Four Main Stages + Exit Criteria, with a Full Set of Copy-Paste Prompts

A 36-page official playbook broken into graduation criteria, common traps, and matching Claude prompts for four stages — ready to copy and use
One-Minute Brief
  • In May 2026 Anthropic released the 36-page founder's playbook The Founder's Playbook, re-splitting the AI-native startup path into four stages — Idea, MVP, Launch, Scale — each with explicit exit criteria.
  • The core tool stack is the Claude (research & strategy), Claude Code (writing code), Claude Cowork (ops automation) trio; the playbook gives a concrete task-to-tool mapping table.
  • Every stage lists AI-era failure modes: treating a prototype as validation, frictionless scope creep, AI-amplified confirmation bias, agentic tech debt — each with matching Claude exercises.
  • MVP exit criteria are quantified with the Sean Ellis test: if more than 40% of active users say they would be "very disappointed" if they could no longer use the product, treat that as a product-market fit signal.
  • The appendix profiles a dozen real startups, including Carta Healthcare — 22,000 surgical cases a year, 66% faster data abstraction — and Anything, which has helped 1.5 million users turn ideas into software.
Stance note: This is Anthropic's official founder's playbook; the recommended stack throughout — Claude, Claude Code, Claude Cowork — is its own product suite. Efficiency numbers in the case studies (e.g. 66% improvement, 1.5M users) are company self-reports. The four-stage framework and exit criteria themselves can be used on their own.
1What this playbook teaches

An official playbook that redraws the startup path as four stages

On May 6, 2026, Anthropic released a 36-page founder's playbook, The Founder's Playbook (PDF v3), re-breaking the AI-native startup path into four stages: Idea, MVP, Launch, Scale.

The core argument is blunt: what founders must do has not changed — still those three things: find a real problem, build something that solves it, turn it into a company. AI changes the road to the finish line. It compresses that path so each stage no longer automatically demands a larger team and another funding round; the bottleneck moves from "what you can build" to "what you choose to build."

What sets it apart from generic startup pep talks: from idea validation all the way to scaled exit, every stage's "when to move on" is written as checkable decision conditions, not principle-level advice — and behind each condition hangs a set of Claude prompts you can copy as-is.
1Idea 2MVP 3Launch 4Scale
Idea
  • All three judgment questions answer "yes"
  • Problem-solution fit found
MVP
  • Evidence on at least one of retain / pay / refer
  • Sean Ellis test above 40%
Launch
  • Growth is repeatable and channel-driven
  • Product holds production load
  • Ops no longer bottlenecked on the founder
Scale
  • One of three threshold events
  • Sustainable profit / IPO-ready / acquired

The playbook keeps returning to the same tool trio: research and strategy to Claude, code to Claude Code (an environment where AI writes, tests, and debugs code autonomously), ops automation to Claude Cowork. Below we break it stage by stage — first, which part of the path AI actually compresses.

2Where the compression happens

The old path was a seven-step loop; the new path keeps four main stages

AI does not let startups skip any stage. What it unhooks is the default binding between stages: every new stage used to mean hire more people, swap skill sets, raise another round. That binding is now undone.

Traditional path · every step triggers expansion
ValidateFundraiseHireBuildRaise againGrowHire more↻ loop
AI-native path · stages no longer auto-expand team / capital
IdeaMVPLaunchScale

The playbook is concrete about what AI replaces at each layer: in research, it is an on-call expert across domains — competitive analysis, market sizing, financial modeling, drafting investment memos; in coding, it is an always-online, never-blocked engineer that compresses what once took a whole team into work a founder can ship alone; in ops, it is an on-demand automated ops team — CRM updates as deals advance, weekly reports write themselves, product docs stay in sync as the product changes.

The result is that the founder's center of gravity moves up: from "executor" writing code, managing people, and handling daily chores, to "orchestrator" generating ideas and directing these AI tools to land them. The playbook flags the highest-leverage change: people with domain experience but no engineering background can start companies too — the founder pool expands beyond engineers.

3First, pick the right surface

Same Claude, three different workspaces

The playbook stresses: Chat, Claude Cowork, and Claude Code are the same Claude underneath; only the workspace wrapped around it differs. Pick the wrong surface and work that should feel smooth turns awkward. The table below is a selection framework you can use as-is.

If the task isUseWhy
One question, one rewrite, a quick brainstormChatFast, conversational, no setup
Research, analysis, or producing a finished document from your files and systemsClaude CoworkReads folders, connectors, skills, scheduled runs
Writing code, testing, shipping softwareClaude CodeReads the codebase, has diffs, git, a dev environment

The playbook's rule of thumb: Chat handles the constant small stuff of running a company — pull a one-line conclusion from a dense investment memo, double-check a claim before a board meeting; Claude Cowork handles real knowledge work that eats time — turn a folder of customer call notes into a theme-findings doc, every Monday morning pull metrics from connected tools into a KPI brief; Claude Code is the engineers' coding environment — from prototype to production, migrating MVP-era legacy code, without waiting for more hands.

📋 How to choose Chat / Claude Cowork / Claude Code
Task is quick Q&A, rewrite, brainstorm → use Chat (fast, conversational, no setup). Task is research, analysis, or producing a finished document from your files and systems → use Claude Cowork (reads folders, connectors, skills, can run on a schedule). Task is writing code, testing, or shipping software → use Claude Code (reads the codebase directly, has diffs, git, a dev environment). All three are the same Claude underneath; only the outer workspace differs.
4Stage 1 · Idea

Don't rush to write code — validate "is this a real problem" thoroughly

Every founder starts in the same place: a problem they cannot stop thinking about. The Idea-stage mantra is one line — no building until the evidence is in. The work of this stage is research, customer interviews, competitive analysis, and an honest look at evidence against you, all before Claude Code writes the first line of production code.

The playbook sequences Idea as questions to answer in order: Is this problem real, specific, and frequent enough to build a company around? Who hits it, and does that constitute a market? Is anyone else solving it, and how well? What would a solution have to do to truly solve it, and does your idea do that? Together they answer one ultimate question: is this worth building?

The key is get specific before you move. "Expense reports are a pain for everyone" is only an observation — untestable. Convert it to a testable hypothesis and the claim can stand.

✕ One observation (untestable)

"Expense reports are a pain for everyone."

✓ One testable hypothesis

"Finance managers at mid-size companies spend 4+ hours a week reconciling expenses because existing tools don't connect to their accounting software."

Exit criteria: all three questions must answer "yes"

Idea exits when you find problem-solution fit — solid qualitative evidence first (mostly from real conversations) that you are solving a real problem for real people, and only then build the solution. All three questions must be "yes" before you leave:

Is the problem real and specific?
  • You can say who hits it, how often, how severe, and how they cope today
Does the solution aim at the real problem?
  • The one validation revealed — not necessarily the one you first assumed
Is the signal enough to start building?
  • This stage never delivers certainty; waiting for certainty is itself a failure mode — but signal must be strong enough for a rational MVP decision, not mere belief