Claude Harness
TIANTIAN-ZHAN/claude-harness
Scaffolding de agentes IA confiáveis em 5 min — harness engineering pronto
Ver no GitHubKit de scaffolding para construir harnesses de agentes IA em produção em 5 minutos. Implementa práticas de harness engineering com quatro skills essenciais: bootstrap, review, phase completion e lesson promotion. Baseado em pesquisas do AGENTS.md spec e GitHub para garantir confiabilidade e reduzir débito técnico.
Instalação
npx convez-ai TIANTIAN-ZHAN/claude-harness1 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
- Estabelecer rulebook e convenções em novo projeto com IA
- Revisar e sincronizar documentação com código após milestones
- Rastrear débito técnico e lições aprendidas de forma estruturada
- Criar contexto eficiente para agentes Claude reutilizarem entre sessões
Veredito

Ferramental sólido que endereça um problema real: agentes IA em produção precisam de regras e contexto para serem confiáveis. A abordagem em três estágios (Seed/Standard/Full) é pragmática e os skills são bem desenhados. Recomendado para times que levam harness engineering a sério.
README
A tradução para o português deste README ainda está na fila.
Claude Harness
Scaffold a production AI agent harness in any repo — in 5 minutes.
Tools for the post-vibe-coding era.
In February 2025, Andrej Karpathy coined "vibe coding": "fully give in to the vibes, embrace exponentials, and forget that the code even exists." By December, Collins named it Word of the Year. By Q1 2026, Karpathy himself called it passé.
The data caught up. 96% of developers don't trust AI-generated code. 45% of it ships with security flaws. 16 of 18 CTOs reported production disasters. Salesforce is calling 2026 the year of technical debt.
OpenAI and Martin Fowler have a name for what comes next: harness engineering — the rules, context, and feedback loops that make agents reliable.
This is the toolkit. Four skills, five minutes per repo.
The four skills
| Skill | Treats | When |
|---|---|---|
/bootstrap-harness | A repo with no rulebook | Fresh project · no CLAUDE.md · major refactor |
/review-harness | Docs that drift from code | After a milestone · monthly |
/finish-phase | Sprints that never close out | A milestone ships |
/promote-lesson | The same mistake, twice | Same lesson referenced ≥ 2 times |
Install
/plugin marketplace add TIANTIAN-ZHAN/claude-harness
/plugin install claude-harness@TIANTIAN-ZHAN
Or copy into ~/.claude/skills/:
git clone https://github.com/TIANTIAN-ZHAN/claude-harness ~/.claude/skills-tmp
cp -r ~/.claude/skills-tmp/skills/* ~/.claude/skills/
What /bootstrap-harness builds
A docs system the agent reads before doing anything:
CLAUDE.md— commands, conventions, three-tier boundaries (✅ Always / ⚠️ Ask first / 🚫 Never). Capped at 150 lines so it stays in context cheaply.docs/exec-plans/—active/for in-flight work,completed/for shipped. The index file becomes the new-session entry point: open it and the agent knows where you left off.docs/lessons.md— append-only post-mortems. The raw material/promote-lessonmines.docs/exec-plans/tech-debt-tracker.md— the "we'll come back to this" list, prioritized P0–P3.- Optionally:
GLOSSARY.md,ARCHITECTURE.md,core-beliefs.md— only when the project earns them.
Three progressive stages — Seed for any repo (~5 minutes), Standard for MVP-level (+30 minutes), Full for mature systems (+2 hours). Stop wherever fits your project's maturity.
How it's built
Encodes three pieces of research from the past year — the AGENTS.md spec (stewarded by the Linux Foundation, used by 60,000+ projects), GitHub's 2,500-repo study on what AGENTS.md / CLAUDE.md files actually work in the wild, and the ACE paper on structured-bullet + delta-edit patterns for evolving rulebooks (the basis for /promote-lesson).
MIT licensed.