project-compound
/project-compound
Zapisuje i wyszukuje wiedzę projektową w katalogu docs/agents/knowledge po naprawach błędów, decyzjach projektowych i pracy w nowych modułach.
- Autor
- guangbin79 ↗
- Repozytorium
- silicon-memory ↗
- Licencja
- MIT
- Ostatnia zmiana repozytorium
- 13.11.2026
Ocena i bezpieczeństwo
- Pliki w paczce
- 13
- Skrypty
- 3
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 0
Brak sygnałów. Repozytorium zmienione 13.11.2026, bez odwołań do wycofanych modeli i API.
MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.
- 3 skrypty w paczce. Przejrzyj je przed instalacją, bo Claude może je uruchomić.
Skrypty: evals/fixtures/scripts/pc-check-decision-entry.sh, evals/fixtures/scripts/pc-check-feedback.sh, evals/fixtures/scripts/pc-check-no-material.sh
Co robi i kiedy
Zapisuje i wyszukuje wiedzę projektową w katalogu docs/agents/knowledge po naprawach błędów, decyzjach projektowych i pracy w nowych modułach.
Oryginalny opis (EN)
Use when the user says 知识积累/知识沉淀/记录知识/知识更新/知识查询 (knowledge accumulation/recording/update/query), or when starting a task on an unfamiliar module, after fixing bugs, after making design decisions, when investigating issues across modules, or when project knowledge needs recording or querying from ./docs/agents/knowledge/.
Jak działa
- Overview
- When to Use
- When NOT to Use
- Three Operations
- Directory Structure
- Bootstrap (first ingest only)
- References (read on trigger, not upfront)
- Operation: Ingest (fast path)
- Operation: Query (fast path)
- Operation: Lint (fast path)
Podgląd SKILL.md
---
name: project-compound
description: |
Use when the user says 知识积累/知识沉淀/记录知识/知识更新/知识查询 (knowledge accumulation/recording/update/query), or when starting a task on an unfamiliar module, after fixing bugs, after making design decisions, when investigating issues across modules, or when project knowledge needs recording or querying from ./docs/agents/knowledge/.
---
> ⚠️ Superseded by `unified-memory`(sidecar Phase 3,2026-10-11):memory+knowledge 已有统一入口;本 skill 语义与数据面不变,references 仍是唯一真源。
> 统一入口:`skills/unified-memory/SKILL.md` | CLI:`silicon-memory knowledge {query|ingest|lint}`(机械面)+本 skill Ingest/Lint 过程(语义面)
> 本 skill 通常由 `settle` 返回的 `agent_actions`(ingest-if-material)触发;下列命令为内部执行器
# Project Compound
## Overview
Incremental, compounding knowledge wiki for this project. The LLM builds and maintains a structured wiki that sits between raw experience and future problem-solving. Knowledge is compiled once, cross-referenced, and kept current — not re-derived on every task.
**Core principle:** The wiki is a persistent, compounding artifact. Every completed task makes it richer. Cross-references are already there. Contradictions are flagged. The synthesis already reflects everything learned so far.
## When to Use
- Task completed (feature, bugfix, refactor)
- Iteration landed or worktree about to be torn down (ingest before cleanup — knowledge outlives the worktree)
- Design decision with non-obvious tradeoffs
- Bug root cause identified
- Need to query prior knowledge before acting
- Periodic knowledge health check (lint)
- All tasks in tasks.md completed (trigger archive)
## When NOT to Use
- Trivial changes (typo, single-line) — don't create module entries; if a closing ingest is mandated anyway, walk the flow and take the No material log-only path
- Task still in progress
- Obvious rationale, no learning value
## Three Operations
| Operation | Purpose | Trigger |
|-----------|---------|---------|
| **Ingest** | Integrate new knowledge into wiki | Task completed, decision made, bug fixed |
| **Query** | Search wiki, synthesize answer, file back | Before acting on unfamiliar module, investigating issue |
| **Lint** | Health-check wiki for drift and gaps | Periodically, or before major work on existing module |
## Directory Structure
```
docs/agents/knowledge/
knowledge-index.md # Content-oriented catalog (Layer 1)
log.md # Chronological append-only log
feedback.md # Entry application feedback (append-only, cross-task)
modules/ # Module knowledge pages (Layer 2)
{module-name}.md # Interlinked, structured per template
```
## Bootstrap (first ingest only)
Before the first ingest, create only what is missing — never overwrite existing files:
- `modules/` directory
- `knowledge-index.md` with the seed table:
```markdown
# Knowledge Index
## 模块列表
| Module | Description | Summary Path | Detail Directory |
|--------|-------------|--------------|------------------|
```
- `log.md` with heading `# Knowledge Log`
Files already exist? Skip silently.
## References (read on trigger, not upfront)
- Ingest 收尾执行/条目四类型模板/写作规则不熟 → read `references/operations.md`(Step 0-6 全语义)
- Lint 执行或判据存疑 → read `references/lint.md`(Tier 1/2/3 全表)
- 写出的内容疑似坏味道 → read `references/anti-patterns.md`(Common Mistakes)
常规操作走下方快路径即可,不必通读 references。
## Operation: Ingest (fast path)
0. **应用回写(先于分类)**:本次任务实际应用了既有条目 → 向 `feedback.md` 追加一行 `{YYYY-MM-DD} | {module}/{slug} | {project} | {harness} | worked|needed-adjustment|obsolete | {一句话说明}`(条目用其 `**Id:**` 引用;仅真实应用才写;语义细节见 references/operations.md Step 0)
1. **Classify**:Chose between alternatives → Decision;non-obvious approach → Strategy;found and fixed a bug → Bug Experience;failed approach not to repeat → Failed Approach;module structure changed → Module Info;task plan completed → Task Archive。None match → **No material**:log-only entry(Types: No material),skip Steps 2-5;never force entries
2. **Target module**:component X → X;cross-cutting concern (auth, error handling) → topic name;infrastructure (build, CI) → `infra`;documentation → file topic (e.g. `readme`);uncertain → check `knowledge-index.md`
3. **Write record**:create/append `modules/{module-name}.md` per the full template in `references/operations.md`(每条目首 bullets:`**Id:** {module}/{slug}` + `**Source:** {commit hash | ./tasks/{file} | session id}`;Decision 另含 Chosen/Alternatives/Reason/Tradeoff/Supersedes)
4. **Cross-reference**:decision affects other modules → add `See also` in those modules;fix relates to prior bug elsewhere → link both;new dependency → update both Overviews
5. **Update index**(new module only):add row to 模块列表 — module name, one-line description, summary path, detail directory
6. **Append to log**(append-only, parseable):
```markdown
## [{date}] ingest | {module-name}
- {1-2 line summary of what was recorded}
- Types: {Decision/Strategy/Bug/Failed Approach/Feedback/Module Info/No material}
```
`grep "^## \[" log.md | tail -5` shows last 5 entries.
## Operation: Query (fast path)
Read `knowledge-index.md` → read relevant module pages → synthesize answer → return.
Cross-project borrowing: before porting any rule/decision/config from another project, run query (and ingest for new material) against THAT project's `./docs/agents/knowledge` first; copy conclusions with their evidence, never blindly.
Good answers compound. If a query produces a comparison, analysis, or connection not yet in the wiki — file it back:
1. Append the synthesis to `./docs/agents/knowledge/modules/{module}.md` as a new entry(套最接近的条目模板;`Source: query synthesis` + 日期)
2. Append to `log.md`:
```markdown
## [{date}] query | synthesis -> {module}
- Types: Synthesis
```
Likewise: if this task actually applied an existing wiki entry, write one feedback.md line at the closing ingest (Ingest Step 0). Only on actual use - never as ritual.
## Operation: Lint (fast path)
- **Tier 1 safe fixes(auto-apply)**:modules/ file missing from index → add row with `(no summary)` placeholder;index row points to nonexistent file → prefix Description with `[MISSING]`,never delete the row;index/module drift → page wins, sync index
- **Tier 2 decision issues / Tier 3 report-only**(broken links、asymmetric supersede、dead Source links、contradictions、stale claims、orphan/missing pages、resolved open questions、gaps)→ report, never auto-fix;full tables in `references/lint.md`
- Append to log.md: `## [{date}] lint | {N} issues found, {M} auto-fixed` plus one bullet per Tier 1/2 finding
## Quick Reference
| Action | How |
|--------|-----|
| Ingest after task | Classify → Write → Cross-ref → Index → Log |
| Query knowledge | Read index → Read modules → Synthesize → File back if new |
| Lint wiki | Tier 1 auto-fix → Tier 2/3 report → log `N issues found, M auto-fixed` |
| Add cross-reference | `See also: [module-name](module-name.md)` in Overview |
| Supersede old decision | Add **Supersedes** in new, mark old with `> **Superseded by:**` |
| Check recent activity | `grep "^## \[" ./docs/agents/knowledge/log.md \| tail -5` |
| Archive completed task | Read ./tasks/{file}, run ingest, mark in ./tasks.md as [DONE] |
| Write entry feedback | Task applied an entry -> Ingest Step 0: append feedback.md line `date \| module/slug \| project \| harness \| outcome \| note` |
| First ingest | Bootstrap: create missing index/log/modules seeds |
## Common Mistakes
全表(12 条 Mistake|Fix)见 `references/anti-patterns.md`。高频三条:one fact per bullet(禁段落);`./` 相对路径;琐碎变更走 No material log-only 路径。
## Related Docs
- `./docs/agents/knowledge/knowledge-index.md` - content index
- `./docs/agents/knowledge/log.md` - chronological log
- `./docs/agents/writing-guide.md` - documentation conventions
- `./docs/agents/index.md` - overall docs structure
Pliki w paczce (13)
- SKILL.md 8 KB
- evals/eval.yaml 1 KB
- evals/cases/pc-ingest-bootstrap.yaml 1 KB
- evals/cases/pc-ingest-feedback.yaml 2 KB
- evals/cases/pc-ingest-no-material.yaml 1 KB
- evals/cases/pc-lint.yaml 2 KB
- evals/cases/pc-query.yaml 1 KB
- evals/fixtures/scripts/pc-check-decision-entry.sh 1 KB
- evals/fixtures/scripts/pc-check-feedback.sh 1 KB
- evals/fixtures/scripts/pc-check-no-material.sh 1 KB
- references/anti-patterns.md 1 KB
- references/lint.md 2 KB
- references/operations.md 6 KB
Wersje
W bazie nie ma innych skilli o nazwie project-compound.
Migawka bazy z 13.11.2026 · wersja paczki cf73906dfd6f (skrót SHA-256 całej paczki)
Pobrania w ostatnich 30 dniach: 0 · ranking
Zgłoś problem z tym skillem: szkodliwa treść, błędny opis albo prośba autora o usunięcie.
Podobne skille
- 90unified-memory
Przeszukuje i zapisuje wspólną pamięć agenta, obejmującą doświadczenia z projektów i strony wiedzy, jednym interfejsem. - 100compound-docs
Zapisuje rozwiązane problemy w projektach Elixir, Phoenix i Ecto jako przeszukiwalną dokumentację z frontmatterem YAML, budując wiedzę zespołu. - 100project-knowledge
Generuje wiedzę o repozytorium git: graf kodu, wiki i opisy domeny w katalogu .ai, przy wdrażaniu w projekt lub odświeżaniu nieaktualnych danych. - 100compound-docs
Tworzy przeszukiwalną bazę rozwiązań dla projektów Elixir, Phoenix i Ecto z frontmatterem YAML i wykorzystuje ją przy diagnozowaniu nowych problemów. - 85project-engineering
Prowadzi badanie i wdrażanie zmian w projekcie oraz dba o wiedzę inżynierską projektu w trakcie pracy. - 90distill-memories
Przenosi trwałą wiedzę z notatek pamięci długoterminowej do głównych dokumentów projektu: przeglądu i architektury.