Claude's Founder's Playbook, and the agent skill I built from it

Anthropic published a 36-page guide called The Founder’s Playbook: Building an AI-Native Startup. I read it as a founder who builds with Claude every day, and it’s the most honest description I’ve found of how company building works now, when AI writes the code and the founder decides what’s worth building.

The playbook is free and you can read it here. This post is what stuck with me, plus what I did with it after: I turned the whole thing into a skill my agents can run.

What’s in the playbook?

It remaps the four stages of a startup, Idea, MVP, Launch and Scale, for founders using AI as core infrastructure. Each stage gets a goal, exit criteria you can test yourself against, the failure modes that show up at that stage, and concrete ways to apply Claude at each step.

The framing I keep coming back to: the founder’s job hasn’t changed. Find a real problem, build something that solves it, scale it into a company that matters. What changed is the path, because AI compresses quarters into weeks and the founder’s role shifts from doing the work to directing the systems that do the work. The line that sums up the whole document: “The bottlenecks are no longer what you can build, but what you choose to build.”

The four ideas that stuck with me

Building is no longer validating. The playbook cites the stat that 42% of startups fail because they built something nobody wanted, and it argues that number is going to climb, because agentic coding collapsed the distance between “I have an idea” and “I have a product.” A working prototype feels like evidence, but what it actually gives you is a prop for conversations with real users, and those conversations are where the evidence comes from.

AI technical debt compounds differently. Some shortcuts are fine at MVP. The new failure mode is building without specs and architectural decisions written down somewhere the AI can read, like a CLAUDE.md file, so every session re-derives your decisions from scratch and they drift. You end up with a codebase that has no coherent mental model behind it, not because any single piece is bad, but because the pieces were never designed to fit together.

Launch energy is not product-market fit. Your friends, a Hacker News spike, warm intros from your investor’s portfolio. The playbook’s point is that none of that predicts what week twelve looks like, so it tells you to set retention benchmarks before you release anything, and to define in advance what a false positive would look like for your product.

Confirmation bias now comes with a research engine. Ask AI for evidence supporting what you already believe and it will find it. The antidote is the same tool pointed the other way: ask Claude to refute your idea, argue the strongest competitor’s case, run the pre-mortem. The playbook calls this a core use case at every stage, and it matches my experience.

Where I saw my own mistakes in it

I’m building my company solo, with agents automating a lot of the daily work on my own hardware, a mac mini for now. I’ve also built products nobody wanted, more than once. The building was always the fun part, the validating was the part I skipped, and the playbook names that exact trap on page 10, so reading that chapter felt uncomfortably familiar.

A PDF is written for humans

Here’s my problem with great playbooks: I read them, I agree with them, and three weeks later I’m making decisions from memory again. And if agents do the daily work now, the advice shouldn’t live in my head anyway. It should live where the agents can reach it in the middle of a decision.

So I distilled the playbook into an agent skill. A skill is a portable folder of instructions that any skills-compatible agent can load: Claude Code, Cursor, Codex, Gemini CLI. When your task matches the skill’s description, the agent pulls it in on its own, with no prompting and no pasting.

The founder-playbook skill is the lifecycle layer. It diagnoses which stage you’re in, names the goal and the failure mode for that stage, and points at the single most important next move. Ask your agent “am I ready to launch?” and instead of general encouragement, you get the playbook’s actual exit criteria applied to your situation.

Thirteen skills, one command

I packaged the hub with 12 per-function skills that carry the depth: idea validation, product, business model, go-to-market, growth analytics, customer success, sales, marketing, fundraising, finance, operations, and legal. Those 12 are adapted from Linas Beliūnas’s One-Person Unicorn set, and the hub routes to them when you need a concrete deliverable instead of direction. Credit where it’s due: the frameworks are Anthropic’s and Linas’s, and the packaging into portable, spec-conformant skills is mine.

Install the whole collection:

npx skills add samuelcastro/startup-skills

Or just the hub skill:

npx skills add https://github.com/samuelcastro/startup-skills --skill founder-playbook

Everything is on GitHub, MIT-licensed, and each skill follows the open Agent Skills spec so it works across tools instead of being locked to one.

Does it actually help?

I don’t know yet whether having the playbook in the loop changes my decisions or just makes me more confident in the ones I was already going to make. I’m testing it on my own projects first, and I’ll write about what I find.

The playbook says the bottleneck is what you choose to build. I chose to build the thing that remembers the playbook for me.