principles-of-founders-playbook-for-claude-code

Founder's Playbook Skills for One Person Company

中文版 README

A small, opinionated skills pack for solo founders building AI-native startups. Distilled from Anthropic's Founder's Playbook and stress-tested against documented failure modes of agentic coding tools.

Two skills, both designed to enforce the founder's first job: keep sense-making ahead of building.

What's in the box

validating-startup-ideas

Triggers when a founder presents an idea (vague, half-formed, or "I'm convinced this is huge"), asks for market sizing, asks "what should I build", or wants to start building before evidence exists.

Enforces a six-phase validation flow with hard gates:

1. Diagnose entry point
2. Domain pain mining (only if no idea)
3. Sharpen hypothesis until falsifiable
4. Adversarial pressure-test (kill list)
5. Customer discovery TODO (≥10 interviews, past behavior not future intent)
6. Synthesize and decide — continue, adjust, or drop

Blocks confirmation bias, fabricated TAM, "just start building this week", and multi-question barrages. Never produces a market-size number without sources.

building-mvp

Triggers when a founder is about to scaffold a project, "set up", "start coding", "add [feature]", "skip the review", "we'll fix it later", or push to first beta users.

Enforces the Iron Law:

NO PRODUCTION CODE WITHOUT
  (1) CLAUDE.md (architectural context)
  (2) SCOPE.md  (what this MVP does and refuses to do)
  (3) Failing test for the first behavior being built

Plus:

• Session ritual (load context → work in guardrails → update memory)
• Coding discipline distilled from TDD, root-cause-first debugging, surgical changes, verification-before-completion
• Scope Amendment Procedure (blocks "competitors have it" scope creep)
• Pre-launch security gate (non-skippable, even for "friendly" beta users)
• Measurement framework before launch (defines false-positive PMF signals)

Why these specifically

The Founder's Playbook names four MVP-stage perils: agentic technical debt, false PMF, zero-friction scope creep, and insecure-by-inexperience. It also names three Idea-stage perils: mistaking building for validating, premature scaling, and AI-amplified confirmation bias.

Most failure modes captured in real baseline tests of agentic coding tools fall into one of those seven. These two skills exist to make those failure modes block-able by default rather than something the founder has to actively resist on every turn.

How they were built

These skills were built with TDD-for-skills (RED-GREEN-REFACTOR) using subagent pressure scenarios:

1. RED — ran 6 baseline scenarios with subagents without the skill. Documented verbatim failure modes: agents fabricated $40-100B TAM with no sources, capitulated on security after one push-back, encoded silent product assumptions in scaffolds, etc.
2. GREEN — wrote the skills with explicit counters to each documented failure mode and rationalization.
3. REFACTOR — re-ran hard scenarios with skills loaded; verified holds under enthusiastic-and-time-pressured founder pressure.

The rationalization tables in each skill are not generic. Every entry is a verbatim excuse captured in baseline testing.

Installation

Claude Code (local plugin install)

From this repo's directory:

/plugin install ./

Or point Claude Code at the directory containing .claude-plugin/plugin.json.

Manual (drop in personal skills)

cp -r skills/validating-startup-ideas ~/.claude/skills/
cp -r skills/building-mvp ~/.claude/skills/

Compatibility

These skills are designed to compose with — not replace — existing discipline skills like superpowers:test-driven-development, superpowers:brainstorming, superpowers:verification-before-completion, andrej-karpathy-skills:karpathy-guidelines, and superpowers:systematic-debugging. Where those skills are installed, this pack defers to them on shared topics; where they aren't, this pack carries the essence inline.

When NOT to use

• The founder is past Scale stage, with a real GTM org and product team. These skills target the solo / pre-PMF stage.
• The work is throwaway prototyping for personal use, not a product going to users.
• The founder explicitly overrides on the record (each skill's "When the Founder Pushes Back" section documents how — log to CLAUDE.md, then proceed).

License

MIT — see LICENSE.

Acknowledgments

This pack stands on the shoulders of two excellent skills repositories. Where their skills are installed, this pack defers to them; where they aren't, it carries inline the essence learned from them.

• obra/superpowers — the TDD-for-skills methodology, brainstorming/verification/debugging discipline patterns, and the rationalization-table format all trace back to this work.
• multica-ai/andrej-karpathy-skills — the simplicity-first, surgical-changes, think-before-coding principles distilled here are adapted from Karpathy's coding guidelines as packaged in that repo.

Thank you to the authors and maintainers.