Skin in the Game
evanmcgoogan/skin-in-the-game
Contrato operacional que transforma seu agent de yes-man em CTO com skin in the game
Ver no GitHubUm contrato operacional explícito para agentes Claude Code que define clara divisão de responsabilidades: o agente toma decisões técnicas e reversíveis com autonomia, escalando apenas questões de estratégia, gosto e one-way doors. Implementa 12 princípios com skills que forçam autoavaliação rigorosa antes de qualquer entrega.
Instalação
npx convez-ai evanmcgoogan/skin-in-the-game1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Evitar builds confiantes de ideias mal-formadas
- Escalar apenas decisões estratégicas genuinamente suas
- Impor autoavaliação rigorosa antes de llamar código 'pronto'
- Questionar e deletar requisitos desnecessários sistematicamente
Veredito

Ferramenta sofisticada que resolve um problema real: agentes que ou concordam com tudo ou pedem permissão demais. O contrato operacional é bem pensado e a execução com skills compostos é sólida. Requer disciplina de uso, mas transforma radicalmente a dinâmica agent-human em projetos de código.
README
A tradução para o português deste README ainda está na fila.
🤝 Skin in the Game
The Operating Contract. Give AI agents Skin in the Game.
Twelve build principles that turn your coding agent from a yes-man into a COO/CTO with skin in the game — it owns the technical calls, escalates only what's truly yours, and grills you before it builds the wrong thing.
MIT licensed · for Claude Code
The problem
Your coding agent has two failure modes, and you've hit both:
- The yes-man — it agrees with your half-formed idea and confidently builds it, flaws and all.
- The runaway — it asks permission for trivia, or charges through a one-way door it should have flagged.
The fix isn't a cleverer prompt. It's a contract — an explicit division of labor that tells the agent what to own and what to bring to you.
Before / after
Without a contract:
You: Add caching to the API.
Agent: Great idea! Adding Redis now. …and it ships a cache, a per-route config layer, and a warming cron. It never asked whether the data can go stale. You find out in production.
With Skin in the Game:
You: Add caching to the API.
Agent: Owning the technical calls: in-process LRU keyed on tenant + query — we're on one box, so Redis is a dependency we don't need yet. Cutting the per-route config — flexibility with no second use case (#7, delete the requirement).
One decision is genuinely yours, because acceptable staleness is a one-way door for what users see: (a) 60 seconds — safe everywhere, dashboards lag a beat (my recommendation) · (b) 5 minutes — cheaper, but billing pages can show stale numbers.
Everything else proceeds now. Constitution-check runs before I call it done.
Same request. One question reaches you — the one that was actually yours.
Install
/plugin marketplace add evanmcgoogan/skin-in-the-game
/plugin install skin-in-the-game@skin-in-the-game
Start a new session and the contract is in context from turn one. It re-loads after every /clear and compaction, so it never silently falls out.
What you get
This isn't a CLAUDE.md you paste once and forget. It's a methodology — twelve principles plus six composable skills and a session-start hook that enforce them:
| Skill | What it enforces |
|---|---|
| operating-contract | #0 — the spine: own technical/reversible calls, escalate only taste/strategy/one-way-doors |
| grill-me | #6 — sharp multiple-choice questions at the CEO line, before building |
| delete-the-requirement | #7 — question every requirement, then try to delete it |
| research-first | #8 — read before you write; verify before "done" |
| reality-check | #4 & #5 — kill fake passes, vanity metrics, and performed rigor |
| constitution-check | the gate — self-review any plan or diff against all twelve |
And before the agent calls any substantial build done, it grades its own diff:
Constitution Check — cache layer, before calling it done
✅ Honored: P8 read the existing fetch path first · P9 reused the repo's LRU helper
⚠️ Violations:
- P7 Delete: added a cache-bypass flag "for debugging" — nobody asked for it → cut
- P5 Honesty: about to report "tests pass" — the staleness path was never exercised
→ run it, or report it unverified
🚪 Escalate: none — staleness was settled up front
How it works — the altitude line
The whole system turns on one line:
Below the strategy/taste line, the agent replaces you. Above it, it amplifies you.
- Below the line (architecture, libraries, naming, anything reversible) → the agent decides and proceeds. No rubber-stamping.
- Above the line (taste, strategy, one-way doors) → it escalates as a decision: plain-English options, pros/cons, a recommendation. Never a status update.
Because you're rate-limited by exactly one resource — your time and attention — and it should be spent only where it's irreplaceable.
The twelve
The full constitution lives in PRINCIPLES.md. In brief:
The North Star — 1. Amplify above the line; own below it · 2. Compete on a different plane · 3. Build for the durable thesis Intellectual Honesty — 4. Measure only what reality grades · 5. Honesty over performance Engineering Discipline — 6. Grill me at the CEO line · 7. Delete the requirement · 8. Research first · 9. Build on top, not next to · 10. Density over breadth · 11. Clean over Frankenstein
(0 · The Operating Contract sits above all eleven.)
Why this is different
Every other framework makes your agent better at doing — TDD, debugging, planning. Skin in the Game governs the relationship — who decides what, when to escalate, how to stay honest. It's the first operating contract for building with an AI agent. That's the category.
License
MIT © evanmcgoogan. Adopt the principles, fork them, make them yours.