Team knowledge, compiled — not just stored. 12 typed templates · 14-step ingestion pipeline · 4-layer retrieval · 3-tier edge trust model. Every answer cites its source, or honestly says "no record".
Quickstart · What it does · Triggers · Layout · When NOT to use · CHANGELOG · 中文版
- Answers from model memory instead of your team's actual decisions — confidently wrong
- Forgot yesterday's postmortem while writing today's code
- Can find a doc, but can't tell you "what breaks if I change X"
- Meanwhile your wiki rots silently — nobody knows which doc is stale
This skill fixes that with one mechanism: documents are the single source of truth, and the knowledge graph is a pure function of them (graph = compile(documents)). Delete .knowledge/.graph/ and you're back to a plain doc library — the graph is layered on, never baked in.
$ /knowledge-wiki ask "what's the print-service timeout?"
回答：打印服务超时阈值为 30s，超时后 fallback 到本地缓存，不直接报错。
来源：.knowledge/ops/print-timeout.md（第 12 行）
录入：@alice，2026-05-12
状态：FRESH（距过期还有 82 天）
被引用：[[order-flow-design]]（1 处）
No record? It says so — and refuses to fill the gap with training data (invariant C1).
For humans — 3 steps:
# 1. Install the skill
npx skills add anthonyhann/knowledge-wiki
# 2. Then, in your AI coding agent (Claude Code / Codex / Cursor …), run:
/knowledge-wiki init
# 3. Record your first piece of knowledge
/knowledge-wiki in The print service times out at 30s, then falls back to local cacheinit checks dependencies (rg/git), creates the .knowledge/ skeleton (10 business dirs + index/log/queue system dirs), and installs 5 git hooks.
For AI agents — paste this to your agent:
Install the knowledge-wiki skill from https://github.com/anthonyhann/knowledge-wiki (e.g.
git cloneit into my skills directory, ornpx skills add anthonyhann/knowledge-wiki), then read its SKILL.md and run/knowledge-wiki initin my project.
| Scenario | You say | It does |
|---|---|---|
| Ingest anything | in <text / URL / file / code> |
14-step pipeline: type inference → per-type template → quality gate → index → backlink queue → audit log |
| Ask with receipts | ask "<question>" |
4-layer retrieval (index pre-filter → ripgrep → fuzzy → semantic); 5-element answer with source/owner/freshness/ref-chain |
| Ask relationships | ask graph "X" · ask explain "X" · ask path "A" "B" |
Neighborhood subgraph, full node explanation, multi-hop knowledge chain — authoritative edges only, inferred edges labeled |
| QA knowledge packs | ask test "<module>" |
Aggregates test env/rules/checklists/pitfalls/test-data into one package for QA agents |
| Code → docs | docs --staged |
Analyzes your diff, asks one suggestion at a time, never blocks the commit |
| Keep it alive | health / rot / scan / lint / distill |
Expiry scans, external-source drift detection, 14 consistency checks, dedup & orphan detection |
| See the whole web | health graph-build → graph export --html |
Compiles docs → graph.json; exports interactive HTML / Obsidian Canvas / GraphML |
| Trust governance | health discover → health edge-review |
AI-proposed edges queue up for human review; approved edges are written back into docs, then re-derived. No auto-approve backdoor |
The skill activates when you say things like:
| When you say… | It runs |
|---|---|
| "record this to the knowledge base" / "把这个记到知识库" | in |
| "does the knowledge base have…?" / "查一下" | ask |
| "how are X and Y connected?" / "what breaks if I change X?" | ask graph / ask path |
| "knowledge base health check" | health |
| "document this commit" | docs |
Numbers you can verify in this repo:
- 12 typed templates with per-type required sections + quality gates — no one-size-fits-all "TL;DR + details"
- 14 rule documents in
references/covering ingestion, retrieval, lint, graph, trust, logs - 12 scripts: 6 PEP 723 graph engines (single dependency
networkx, uv-isolated) + 5 git hooks + 1 dependency checker - 14 consistency checks (L1–L9 + 5 graph/pending-queue lints) run centrally, never blocking your flow
- 8 core invariants (C1–C8) + 9 soft behaviors (E1–E9) — e.g. C1: no source, no write, no answer
- 3-tier edge trust model (EXTRACTED / INFERRED / AMBIGUOUS) with a human-review gate; AMBIGUOUS edges can never become hard conclusions
- Shipped in 9 gated phases (v0.1.0 → v2.1.0), each with a documented acceptance matrix — see CHANGELOG.md
This is not a weekend demo. Between 2026-04 and 2026-10 it ran inside a production fleet of 49 agent skills tracked over 342 scored optimization records (measured by skill-eval-optimizer):
- 67.8 → 90.3 / 100 across 34 optimization rounds — every round scored on the same 8-dimension rubric, a change kept only when the score actually moved up
- One round shrank the skill 404 → 216 lines (35 KB → 21.4 KB, −39% tokens) while raising the score — smaller context, better behavior
- Fleet-wide: of 44 skills with full baselines, 40 improved, average +9.9 points — this repo is what survived that loop
knowledge-wiki/
├── SKILL.md # Entry: command skeleton + C1–C8 invariants + E1–E9 soft behaviors
├── DESIGN.md # L1/L2/L3 cognitive-layer architecture + template-library rationale
├── CHANGELOG.md # Full version history (v0.1.0 → v2.1.0)
├── TODO.md # Roadmap & design-decision memos
├── test-prompts.json # Dry-run eval prompts (7 cases: happy + edge)
└── references/ # Lazy-loaded authority docs (loaded on demand, never up-front)
├── *.md # 14 rule docs: ingestion / ask / index / backlink / log / lint /
│ # health / docs / url / graph-model / graph-build / trust / pending-edges
├── templates/ # 12 type templates + registry + 5 system config templates
└── scripts/ # 6 graph engines (Python, PEP 723) + 5 git hooks + check-deps.sh
What init creates in your project: .knowledge/ with 10 business directories (glossary/ design/ requirements/ flows/ case/ tests/ apis/ db/ ops/ synthesis/) organized by the L1 domain / L2 execution / L3 capability layering — see DESIGN.md.
Honest boundaries — this skill is not for:
- Personal notes / diaries / chat logs / throwaway drafts — it's built for team knowledge tied to a codebase
- Teams wanting a hosted, real-time-collaboration wiki — this is local Markdown + git; there is no server, no web UI, no multiplayer editing
- Anyone unwilling to maintain
sources/statusdiscipline — C1 means the agent refuses to answer when there's no record; that's the point, but it's a habit change - Huge corpora (10k+ docs) needing vector search — retrieval is ripgrep + segmented indexes, deliberately embedding-free and inspectable
Built and dogfooded daily by @anthonyhann on real engineering projects — every feature in the CHANGELOG exists because an actual workflow needed it. If it saves your team a re-derivation of "why did we do it this way?", a ⭐ tells the next team it works.
MIT © anthonyhann — see LICENSE.