Production AI: Guardrails e Skills para Agentes de Código
jimtin/production-ai
Skills para agentes de IA provarem trabalho antes de produção
Ver no GitHubBiblioteca de skills e gates que força agentes de IA a comprovar seu trabalho através de testes, planejamento validado e deploys fail-closed. Implementa padrões de qualidade obrigatórios: planejamento prévio, testes preflight, revisões de cobertura, e portas de produção que o agente não consegue contornar apenas com argumentos.
Instalação
npx convez-ai jimtin/production-ai1 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
- Automatizar revisão de código de agentes IA com gates obrigatórios
- Implementar testes e validação preflight em pipelines de agentes
- Orquestrar deploys seguros com fail-closed e proof burden no agente
- Documentar e reutilizar padrões de qualidade entre repositórios
Veredito

Solução madura e pragmática para quem já roada agentes de IA em produção e precisa de guardrails reais, não apenas prompts. Código extraído de setup daily-driver, com documentação clara de cada failure que cada skill previne.
README
A tradução para o português deste README ainda está na fila.
production-ai
Skills and guardrails that make an AI coding agent prove its work.
This is a working library, not a prompt collection. Every artifact here was extracted from a real daily-driver setup where AI agents plan features, write code, run containerized test gates, review pull requests, and promote releases to production — with humans setting policy and agents carrying the proof burden.
The core idea: coding agents optimize for looks done. Left alone, they will write plausible code, skip the boring verification, and report success. Everything in this repo exists to redefine "done" as proven — with tests, evidence, fail-closed gates, and explicit completion blockers the agent cannot talk its way past.
The library
Each skill has two layers:
skills/<name>/— the clean installable payload loaded by the agent when the skill triggers.docs/skills/<name>.md— the public human guide: why the skill exists, the failure it prevents, and how to adapt it.
| Skill | Kind | Guide | Payload |
|---|---|---|---|
| clarify-before-build | Planning gate | guide | payload |
| feature-design-preflight | Design gate | guide | payload |
| user-action-coverage-review | Coverage gate | guide | payload |
| frontend-design-quality | Quality gate | guide | payload |
| test-readiness-preflight | Validation preflight | guide | payload |
| full-app-review | Review orchestrator | guide | payload |
| codebase-prune-review | Removal gate | guide | payload |
| error-logging-instrumentation | Observability review | guide | payload |
| repo-technical-documentation | Documentation engine | guide | payload |
| nextjs-vercel-analytics | Implementation guide | guide | payload |
| pr-production-gate | Deployment gate | guide | payload |
| security-threat-model | Security gate | guide | payload |
| repo-testing-setup | Foundation gate | guide | payload |
| laptop-currency-maintenance | Ops automation | guide | payload |
The patterns
The skills are instances of a small set of system-level patterns, documented in docs/:
- Skill anatomy — the structure every skill follows, and why each part earns its place.
- The skill graph — skills that invoke skills: orchestrators, specialist gates, and composition rules.
- The workspace constitution — one
AGENTS.mdthat governs every repo in a workspace (template). - PR production gate — a fail-closed, container-only review-and-deploy train owned by your own machine.
- The learning loop — a nightly automation that mines the agent's own sessions and proposes skill improvements, safely.
- Sync and backup — version-controlling your agent configuration with redaction and fail-closed pushes.
Quickstart
Install a skill into Codex CLI:
git clone https://github.com/jimtin/production-ai.git
cd production-ai
scripts/install-skill.sh clarify-before-build
Then ask your agent to plan something substantial — the skill triggers on planning requests, or invoke it directly with $clarify-before-build.
Adopt the workspace constitution:
cp production-ai/templates/AGENTS-workspace-template.md ~/workspace/AGENTS.md
Then edit it down: delete every rule you cannot or will not enforce. A constitution the gates don't back up is worse than no constitution — the agent learns the rules are decorative.
Validate the library (used by CI on every push):
./scripts/validate.sh # structure: frontmatter, references, docs, clean payloads
./scripts/privacy-scan.sh # denylist sweep + gitleaks, fail closed
Portability
These artifacts are written for OpenAI Codex CLI (skills live in ~/.codex/skills/, workspace rules in AGENTS.md, cross-references use $skill-name). The patterns port directly to other agent harnesses:
- Claude Code: skills go to
~/.claude/skills/, workspace rules go inCLAUDE.md, and$skill-namereferences become skill-name mentions or slash commands. The anatomy (frontmatter description as trigger, progressive-disclosure references, scripts with tests) is identical. - Anything else: the gates are prose contracts. Any harness that can load instructions conditionally can run them.
What is deliberately not here
This is the public, sanitized derivative of a private setup. Excluded on purpose:
- Client-specific CI integrations and business-domain skills (a market-research/ROI reporting skill, client API contracts).
- Cached data of any kind — API responses, market data, databases, fixtures with real-world records.
- Session logs, learning-loop summaries, and automation state — even sanitized ones.
- Real channel IDs, hostnames, repo names, and absolute paths — replaced with placeholders.
The privacy guarantee is structural, not a one-time cleanup: CI runs a fail-closed denylist sweep plus gitleaks on every push, using the same approach as the private setup it came from. Promotion from the private library into this repo is always a manual, reviewed step — never automated.
Creating content from this repo
The content map tracks which artifacts are ready to become posts, videos, or talks, with the hook for each. The glossary keeps terminology consistent across pieces. Each skill guide is structured to be the first draft of its own article.
License
MIT. Build on it, ship it, write about it — attribution appreciated.