convez.ai
SkillProdutividadeIntermediário
LawlietLuoJia

Curate — Governança de Conhecimento para Claude Code

LawlietLuoJia/claude-code-curate-skill

Audita, deduplicava e promove conhecimento de projeto entre sessões

Ver no GitHub
InicialPluginScore · de 0 a 100

Skill de governança que mantém sua base de conhecimento do projeto limpa e precisa entre sessões. Automatiza deduplicação, atualização, arquivamento e promoção de ativos de conhecimento espalhados em arquivos de memória, CLAUDE.md e documentação.

Instalação

$npx convez-ai LawlietLuoJia/claude-code-curate-skill

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 2 dias
Versão
v2.2.7
Issues abertas
0
Contribuidores
2

Casos de uso

  • Eliminar duplicação de conhecimento em múltiplos arquivos de memória
  • Atualizar e arquivar notas técnicas e decisões obsoletas
  • Mover conhecimento frequentemente referenciado para locais mais visíveis
  • Gerar health scores para ativos de conhecimento (freshness, accuracy, utility)
  • Manter MEMORY.md e arquivos de memória dentro de limites de tamanho

Veredito

Solução robusta e bem-pensada para problema real em projetos longos com Claude Code. O dual mode (Quick/Deep) e feedback loop de scoring são diferenciais sólidos. Recomendado para equipes que usam Claude Code em sprints contínuos.

README

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

Curate — Claude Code Knowledge Governance Skill

中文文档

A knowledge governance skill for Claude Code that keeps your project knowledge base clean, accurate, and maintainable across sessions.

Background

Claude Code uses several mechanisms to persist knowledge across sessions: memory files (under ~/.claude/projects/), project-level CLAUDE.md, and documentation files. Each session can add new memories, update decisions, and modify documentation.

Over time this leads to problems:

  • Duplication: Different sessions save similar information in slightly different wording across multiple memory files
  • Staleness: Decisions and technical notes from weeks ago no longer reflect the current state of the project
  • Bloat: MEMORY.md index files exceed their line limits, individual memories grow beyond recommended sizes
  • Scattered knowledge: The same topic exists as a memory, a CLAUDE.md entry, and a doc — with no coordination between them

None of this is catastrophic, but it degrades the quality of context Claude receives in future sessions. Bad context leads to worse suggestions.

What Curate Does

Curate is a skill you run after a work session (or periodically) that reviews all knowledge assets and applies governance actions:

ActionWhat It Means
DedupFind semantically similar knowledge entries across files, merge them into one
UpdateRefresh entries that no longer match the current project state
RetireRemove or archive entries that are no longer relevant
PromoteMove frequently-referenced knowledge to higher-visibility locations (e.g., memory → CLAUDE.md)
DemoteMove rarely-used knowledge out of high-visibility locations to reduce clutter

Each action is logged with a reason, so you can review what changed and why.

Dual Mode

Quick Mode (5 steps)

Run after any session. Low overhead, focuses on changes from the current conversation:

  1. Size check — Measure MEMORY.md and individual memory sizes against limits
  2. Identify changes — Detect what knowledge was added or modified this session
  3. Govern — Apply dedup, update, retire, promote/demote as needed
  4. Self-check — Verify all edits comply with size and format rules
  5. Audit log — Record each action with rationale

Deep Mode (7 steps)

Run periodically (e.g., weekly or after major milestones). Everything in Quick mode, plus:

  • Health scoring — Each knowledge asset gets a numerical score based on freshness (how recently verified), accuracy (does it match current code), and utility (how often referenced)
  • Deep dedup — Full cross-file scan instead of session-scoped detection
  • Promotion evaluation — Systematic review of which knowledge should move up or down in visibility

The health scoring model improves over time: past governance decisions feed back into the scoring algorithm, so the system learns which types of knowledge tend to become stale quickly and which hold their value.

When to Use

TriggerMode
End of a productive sessionQuick
/curate commandQuick
/curate deep commandDeep
"整理一下" / "curate" / "sync up" / "归档"Quick
Weekly cleanup, post-milestoneDeep

Installation

# Global (available in all projects)
cp -r . ~/.claude/skills/curate/

# Project-level
cp -r . /your-project/.claude/skills/curate/

Project Structure

curate/
├── SKILL.md                  # Skill definition (564 lines)
├── evals/
│   ├── evals.json            # Evaluation criteria
│   └── fixtures/hardening/   # 12 test scenarios
├── references/               # 10 reference docs loaded on demand
│   ├── health-scoring.md     # Scoring algorithm specification
│   ├── pattern-keys.md       # Dedup key definitions
│   ├── evolution-rules.md    # Knowledge evolution rules
│   ├── anti-patterns.md      # Anti-pattern detection
│   ├── knowledge-matrix.md   # Knowledge type taxonomy
│   ├── audit-fields.md       # Audit log field definitions
│   ├── change-matrix.md      # Change detection rules
│   ├── content-matrix.md     # Content type classification
│   ├── deep-output-format.md # Deep mode output format
│   └── governance-insights.md# Governance decision heuristics
├── scripts/
│   ├── test-curate-hardening.js  # Run 12 test scenarios
│   ├── validate-curate.js        # Validate SKILL.md structure
│   └── analyze-curate-history.js # Analyze governance history
└── assets/
    └── governance-insights.md

Test Scenarios

The 12 hardening scenarios in evals/fixtures/hardening/ cover real edge cases encountered during development:

ScenarioWhat It Tests
Quick content changeGovernance after a content-writing session
Quick code changeGovernance after a coding session
Output index onlyOnly output files changed, no knowledge impact
Lightweight noopSession with no meaningful knowledge changes
Deep with known historyDeep mode with existing governance history
Corrupt historyRecovery when governance log is malformed
Missing memoryBootstrap when memory directory doesn't exist
Monorepo CLAUDE.mdMultiple CLAUDE.md files in subdirectories
Large project200+ memory entries, size-aware processing
Anti-pattern cleanupDetecting and fixing knowledge anti-patterns
Cross-project syncDedup across project boundaries
Quick evolution boundaryEdge cases in evolution analysis cutoff

Configuration

Key parameters in SKILL.md:

ParameterDefaultPurpose
MEMORY.md line limit200Keeps the index readable
Individual memory size≤8KBPrevents single entries from dominating
Trigger phrasesCustomizableBoth English and Chinese

Testing

node scripts/test-curate-hardening.js
node scripts/validate-curate.js

License

MIT

Ver o original no GitHub

Bom para

Relacionados