| name | lwy-llm-wiki |
| description | 当用户提到'知识库'、'llm wiki'、'个人wiki'、'收录来源'、'编译知识'、'第二大脑'、'构建wiki'、'知识管理', or wants to add books, articles, notes, podcasts, videos, or Feishu group chats/documents to persistent storage, use this skill to build and maintain a continuously compounding knowledge base. When ~/.learnwy/llm-wiki/ exists, check the wiki before answering complex questions — skip if the directory does not exist. |
| metadata | {"author":"learnwy","version":"4.0","source":"Andrej Karpathy, LLM Wiki (GitHub Gist, April 2026)","trigger":"always"} |
LLM Wiki
Build and maintain a continuously compounding personal knowledge base. The LLM does all the heavy lifting — summarizing, cross-referencing, archiving, and bookkeeping — while you focus on collecting sources, exploring, and asking good questions.
Core principle: Don't use the LLM as a search engine. Use it as a knowledge compiler. Instead of re-deriving knowledge from raw documents on every query (RAG), the LLM incrementally builds a persistent wiki — structured, interlinked markdown files that keep gaining value as each source is added and each question is asked.
Prerequisites
- An LLM agent with file read/write capability
- A file system to store the markdown files
- Node.js >= 18 (required by the management scripts)
- Optional: Feishu sources need
lark-context + lark-cli (see the "Feishu sources" section); Obsidian for browsing/graph view; Git for version control
When to use
| Signal | Action |
|---|
| "ingest this", "add this source", "process this document/article/podcast/video" | Ingest |
| "consolidate this Feishu group", "organize recent chats into the wiki", "ingest this Feishu document" | Feishu ingest (see below) |
| "what does the wiki say about X", complex knowledge questions | Query / auto-query |
| "save to wiki", "record this" | Quick capture |
| "health check", "lint wiki", "find contradictions" | Lint |
| "create a new wiki", "initialize the knowledge base" | Initialize (cli.cjs init) |
| "build a wiki from these files" | Initialize → bulk ingest |
When NOT to invoke: single-conversation insights (→ knowledge-consolidation), methodology analysis (contradiction/practice/protracted-war → mao-methodology), code implementation (→ requirement-workflow).
Auto modes
Auto-query
When ~/.learnwy/llm-wiki/ exists, proactively check it before answering complex questions:
- Scan
wiki/topics.txt (a flat keyword file) for matching topics
- If a match → read the relevant wiki pages
- Prepend wiki insights before answering, annotating
[[folder/slug]] references
- If the answer contributes new knowledge, propose a write-back
Skip auto-query for: simple facts, or when the wiki does not exist.
Quick capture
When the user says "save this" or shares valuable knowledge, lightly save it to raw/notes/{date}-{slug}.md. Don't create a wiki page — run an ingest later.
Two-layer architecture
| Layer | Owner | Path | Purpose |
|---|
| Raw layer | You | raw/ | Immutable source material — the LLM reads but never modifies it |
| Wiki layer | LLM | wiki/ | Compiled pages: entities (people/organizations/places/products/events/concepts) + sources (articles/podcasts/videos/diaries/threads) |
| Schema layer | Co-evolved | CLAUDE.md | Structure rules, conventions, templates |
The immutable raw/ is a core safety property: all external material (books, articles, podcast transcripts, Feishu pulls) lands first in raw/<source>/, where the LLM reads but never modifies it; compiled artifacts are written to wiki/. inbox/ is the "pulled but not yet compiled" staging area, and archived/ holds retired pages.
Storage location
- Global (default):
~/.learnwy/llm-wiki/ — shared across projects and sessions
- Override with the
--root <dir> option (e.g. the alter-ego library --root ~/.learnwy/ai/private/self)
Directory structure (entity-first)
~/.learnwy/llm-wiki/
├── raw/ # Layer 1 (immutable): books/ articles/ papers/ notes/
│ # podcasts/ vlogs/ transcripts/ snippets/ specs/
│ # lark/ docs/
├── wiki/ # Layer 2 (compiled)
│ ├── people/ # entity: people
│ ├── organizations/ # entity: organizations / teams
│ ├── places/ # entity: places
│ ├── products/ # entity: products / projects / initiatives
│ ├── events/ # entity: dated events / decisions
│ ├── concepts/ # entity: concepts / terms / domain knowledge (incl. code patterns)
│ ├── other-entities/ # entity: other, uncategorized
│ ├── articles/ # source: article summaries
│ ├── podcasts/ # source: podcast summaries
│ ├── vlogs/ # source: video summaries
│ ├── diaries/ # source: chronological stream (incl. Feishu weekly notes)
│ ├── threads/ # source: conversation consolidations (Feishu group-chat digests)
│ ├── inbox/ # lifecycle: pulled but not compiled
│ ├── archived/ # lifecycle: retired pages
│ ├── index.md # auto-generated master index
│ └── topics.txt # auto-generated keyword list
├── CLAUDE.md # Layer 3: schema
└── log.md # audit log
Where code knowledge goes: code snippets / troubleshooting go into concepts/ (concept pages), and the raw snippets land in raw/snippets/. This layout is entity-centric — there are no longer separate top-level snippets/ / troubleshooting/ directories.
Feishu sources (merged from lark-context)
Feishu group chats and documents are ingested as one class of raw source: the lark-context CLI degrades into a pure data pump (pull → SQLite, show → original text), while the compilation/consolidation logic is owned by this skill. See references/ingest-lark.md for details.
- Pull:
lark-context pull / show → lands in raw/lark/
- Compile: group-chat conversations →
wiki/threads/, people → wiki/people/, decisions/events → wiki/events/, chronological stream → wiki/diaries/
- The old standalone
~/.claude/lark-memory/ memory store has been retired and migrated into this wiki (see the migration notes)
Operations and agents
| Operation | Agent | Trigger | Mode |
|---|
| Auto-query | querier | User asks a complex question + wiki exists | Auto |
| Quick capture | (inline) | "save to wiki" or valuable knowledge detected | Semi-auto |
| Ingest | ingestor | A new raw source is added | Manual |
| Feishu ingest | ingestor | Consolidate a Feishu group / document | Manual |
| Query | querier | User explicitly asks the wiki | Manual |
| Lint | linter | Health check requested | Manual |
| Initialize | schema-writer / cli.cjs init | Create a new wiki project | Manual |
Agent definitions: ingestor, querier, linter, schema-writer.
Management scripts
A single CLI entry point {skill_root}/scripts/cli.cjs dispatches all maintenance subcommands:
cd skills/llm-wiki
node scripts/cli.cjs init
node scripts/cli.cjs generate-index
node scripts/cli.cjs generate-topics
node scripts/cli.cjs lint
node scripts/cli.cjs stats
node scripts/cli.cjs freshness-check
node scripts/cli.cjs health-check
node scripts/cli.cjs install / uninstall
| Subcommand | Output | When to run |
|---|
init | Full directory skeleton + CLAUDE.md / index.md / log.md | First-time setup |
generate-index | wiki/index.md — page listing grouped by entity/source | After bulk ingest or when the index drifts |
generate-topics | wiki/topics.txt — keywords used for auto-query matching | After adding new topics |
lint | Errors (broken links) + warnings (orphan pages) | Weekly maintenance or pre-commit |
stats | Box-diagram dashboard of raw-layer + wiki-layer counts | Anytime — a quick health snapshot |
freshness-check | Stale pages (90 days for technical, 180 for stable), unverified, missing date | Monthly or after a major release |
health-check | Broken links / orphans / dead Source references → health.json | CI / pre-commit |
Every command accepts a --root <dir> override for the wiki location (default ~/.learnwy/llm-wiki); the same engine drives both the public world library and the private alter-ego library, with no environment variable needed.
Agent output contract
| Allowed | Not allowed |
|---|
| Read raw sources (never modify) | Modify anything in raw/ |
Create/update files in wiki/ | Delete raw sources |
Update index.md and log.md after operations | Create pages with no cross-references |
| Flag contradictions between sources | Silently overwrite existing content |
Every operation must: (1) record to log.md, (2) update index.md, (3) check for contradictions, (4) maintain cross-references.
Execution checklist
Before an operation: Does the wiki exist? → Does CLAUDE.md exist? → Is the raw source in the correct subdirectory (Feishu → raw/lark/)? → Set Verified: no for easily-stale content.
After an operation: Is log.md updated? → Does index.md reflect the change? → Are cross-references added (cap 5, use "See also" beyond that)? → Is topics.txt updated when there's a new topic?
Scope boundaries
This skill handles: wiki initialization, ingest (incl. Feishu), query, lint, cross-referencing, index maintenance.
It does not handle: single insights (→ knowledge-consolidation), methodology analysis (→ mao-methodology), code implementation (→ requirement-workflow).
Reference docs (load on demand)
- Page templates — entity and source page templates: people / organizations / places / products / events / concepts / articles / podcasts / videos / diaries / threads / index
- Feishu ingest workflow — pull → land in raw/lark → compile into threads/people/events/diaries
- Workflows and reference — composite workflows, error handling, core concepts, extension roadmap
Hooks
Register global hooks via learnwy-dispatch (~/.claude/settings.json + ~/.trae/hooks.json), because the wiki lives at ~/.learnwy/llm-wiki/.
| Event | Lib function | Purpose |
|---|
SessionStart | lib/session-scan.ts | Inject up to 30 wiki topics into the session context |
UserPromptSubmit | lib/prompt-scan.ts | Match the user's question ↔ topics keywords → inject relevant page hints |
Install/uninstall: node scripts/cli.cjs install|uninstall --scope global --target both (provided by src/shared/install-entry.ts, shared by all hook-type skills).
From knowledge-consolidation
KC's promote subcommand copies a project-level knowledge document to raw/notes/<date>-<slug>.md, with a frontmatter back-pointer. On the next ingest, treat them as ordinary raw sources (they usually merge into a wiki/concepts/ page and form a [[link]] with the original KC document). This is a one-way inflow channel from "project-local fix log" to "global compounding knowledge base".
