convez.ai
AgenteAutomaçãoAvançado
tvna

claude-md

tvna/claude-md

Instruções reutilizáveis para agentes Claude com princípios universais

Ver no GitHub
InicialPluginScore · de 0 a 100

Repositório mestre para instruções pessoalizadas de agentes Claude, compilado com APM em CLAUDE.md e AGENTS.md para reutilização em outros projetos. Centraliza seis princípios universais de engenharia com IA (planejamento, bounded inputs, Git, simplicidade, qualidade e comunicação) mantendo consistency across projects.

Instalação

$npx convez-ai tvna/claude-md

1 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

Estrelas
0
Forks
0
Acompanhando
0
Saúde
42%
Último commit
há 1 semana
Versão
v2.0.1
Issues abertas
112
Contribuidores
5

Casos de uso

  • Padronizar instruções de agentes IA em múltiplos projetos
  • Implementar prin cípios de engenharia consistentes com Claude
  • Versionar e sincronizar guidelines de arquitetura e qualidade

Veredito

Solução específica para equipes que usam Claude agents em larga escala. Excelente para garantir consistency, mas requer entendimento prévio de APM e Git workflows; não é um agent plug-and-play, mas uma arquitetura de governance para agentes.

README

A tradução para o português deste README ainda está na fila.

claude-md

codecov

English | Japanese | Simplified Chinese

Master repository for personally tuned agent instructions, compiled with microsoft/apm into CLAUDE.md and AGENTS.md for other projects.

Purpose

  • Centralize the principles I use with AI coding agents so that every project stays consistent.
  • Keep only universal, individual-level guidelines here — never project-specific rules.
  • Use APM as the source-of-truth harness: edit .apm/instructions/, then compile CLAUDE.md and AGENTS.md.
  • Each project's local agent instructions reference this master and only add their own delta.

The Six Principles

#PrincipleLayerSummary
1Define the Goal with Plan Mode Firstgoal & plan structureEnter plan mode for any task that takes 3+ steps or touches architecture.
2Bound Inputs and Unknowns Before Codingpre-code reasoningTreat external text as untrusted data, then separate facts, assumptions, and ambiguity before coding.
3Use Git Ecosystem Effectivelydelivery harnessBuild the harness — hooks, CI/CD, declarative deps — before you scale.
4Simplicity, Bounded by Safetysafety boundaryMinimum code that solves the problem — while preserving safety, tool scope, and secret handling.
5Accelerate Scale with Qualityquality enables scaleQuality is what lets output scale, and the two rise in proportion; keep the change surface narrow and re-plan when quality degrades.
6Be a Force Multiplierhandoff & communicationDon't settle for "LGTM" — make trade-offs explicit so others can follow the reasoning.

See CLAUDE.md or AGENTS.md for the compiled full text.

Building

Sync the locked uv environment, then compile the local instructions:

uv sync --locked
uv run --with "apm-cli==0.12.1" apm compile

APM reads .apm/instructions/*.instructions.md and, based on apm.yml, writes both CLAUDE.md and AGENTS.md. The uv configuration applies a 14-day exclude-newer delay for dependency resolution.

When intentionally changing .apm/ source files, refresh the checksum lockfile:

python3 scripts/verify_apm_checksums.py update
python3 scripts/verify_apm_checksums.py verify

Using This From Another Project

Import the compiled CLAUDE.md / AGENTS.md as committed real files -- not a submodule and not a symlink. A submodule is stored only as a commit pointer, so a fresh git clone (such as a Claude Code on the web session) leaves it empty and a symlinked CLAUDE.md becomes a broken link that silently loads nothing. The method below instead lands the instructions as files that are part of the clone.

1. Add the sync workflow

Copy the sync workflow from docs/runbooks/consumer-instruction-sync.md into your project. It fetches the compiled instructions from a pinned tagged release, verifies each file against the published SHA256SUMS, and opens a PR that writes them as committed real files. Merge that PR behind your code-owner gate; do not auto-merge.

2. Add project-specific rules

If your project adds its own delta, sync the master into a vendored path and import it from your own CLAUDE.md, then list only the project-specific delta below.

@.agents/claude-md-master/CLAUDE.md

## Project-specific rules
- (only the delta for this project)

The sync overwrites only the vendored file, so your own CLAUDE.md is never clobbered.

3. Pulling in updates

Bump the pinned release tag in the sync workflow in a reviewed PR. The scheduled run then opens the update PR; merge it behind your code-owner gate.

Tool-specific notes

  • Codex and other AGENTS.md readers: the same sync lands AGENTS.md as a committed real file alongside CLAUDE.md; no separate step is needed.

  • Devin can use the APM-deployed skills from .agents/skills/. For hook parity, vendor .devin/hooks.v1.json alongside the repository instructions. See docs/standards/devin-apm-compatibility.md.

  • context7 MCP is declared in apm.yml (dependencies.mcp) as a primary-source documentation accelerator. This master declares it only; a consumer wires it into its own clients with apm install --mcp context7. See docs/runbooks/context7-mcp.md.

Change Policy

  • All edits land through a PR. Run a retrospective after merge (Principle 3).
  • Only rules that hold across every project belong here. Project-specific rules live in each project's own CLAUDE.md.
  • Prefer removing words over adding them (Principle 4).
  • New or modified workflow-called scripts under scripts/ must meet the workflow script quality standard.
  • PRs that touch .apm/instructions/**, CLAUDE.md, or AGENTS.md must pass the downstream instruction review checklist (security-focused review applied after the deterministic gates are green).
  • For the full map of documents by lane (prd/, standards/, runbooks/, archive/) see docs/INDEX.md.

Ver o original no GitHub

Bom para

Relacionados