A PM knowledge base that stays correct as it grows. Ingest meetings, analyst reports, and strategy docs into a plain Markdown wiki, then work with it from Claude Code or ChatGPT. Every write follows the same rules: cited sources, snapshots before overwrites, explicit supersession, lint.
Based on Andrej Karpathy's LLM Wiki pattern, tuned for product management: competitive intel, customer notes, strategy, roadmap, AI market tracking.
Most PM work today spreads across Slack threads, meeting transcripts, analyst PDFs, and one-off notes. RAG tools like NotebookLM retrieve from raw sources each query, so knowledge never compounds. Personal wikis fail because the bookkeeping overhead outgrows the value.
This skill gives you the middle path: you curate sources, the agent maintains an interlinked markdown wiki that stays current. Every ingest touches 5-15 pages. Every query cites specific wiki entries. The wiki compounds.
Letting an agent write is the easy part. Keeping months of agent writes
trustworthy is the hard part, and it is where most memory tools stop. Here,
when a fact changes the old page is superseded and its inbound links are
rewritten rather than silently overwritten. Every claim points back to an
immutable source in raw/, every overwrite leaves a snapshot in _archive/,
and lint flags stale pages, broken links, and off-schema tags. The same rules
apply whether the write comes from Claude Code or from ChatGPT through the
wiki-mcp server, which is included in the repo and
installed separately.
- Three-layer architecture: immutable
raw/sources, agent-owned wiki pages, andSCHEMA.mdgoverning structure - Ingest / query / update / lint / archive flows with discipline guardrails
- Supersession with auto-redirect of inbound links
- Crystallize pattern: transcripts become structured decision digests
- Privacy-first: pre-ingest filter, and every page is private (excluded from exports) unless it carries
shareable: true - Bundled wiki-search: ranked keyword, regex, tag and backlink search over your whole wiki, via obsidian-mcp (auto-indexes on startup)
- ChatGPT access via wiki-mcp: a remote MCP server behind Cloudflare Access. It is included in the repo but not installed by the plugin; you run it yourself. Read tools by default; with
WIKI_WRITE=1it adds create, edit, append, archive, supersede, and raw capture, each validated againstSCHEMA.md, snapshotted, and logged the same way Claude Code writes are - Obsidian-compatible: works as a vault out of the box
- Prompt recall: each prompt that names a page, or matches pages closely, lists up to 3 of them in context, so the agent reads the wiki before asking you
- Sub-skills (optional, install as needed):
llm-wiki-brief— daily/weekly briefs, tag digests, coverage briefllm-wiki-research— research sprints, competitive deep dives, stub enrichmentllm-wiki-maintain— the daily-maintenance loop (registry-driven sweep → ingest → brief → tidy → health check), with interactive vs autonomous modes
- MY-INTEGRATIONS.md: learned source routing — auto-populated from actual ingest activity
- Next.js embed path for platform deployments
claude plugin marketplace add anh-chu/llm-wiki-pm
claude plugin install llm-wiki-pm@anh-chu-pluginsnpx skills add anh-chu/llm-wiki-pm --skill llm-wiki-pm -a claude-code -gInstall individual sub-skills the same way:
npx skills add anh-chu/llm-wiki-pm --skill llm-wiki-research -a claude-code -g
npx skills add anh-chu/llm-wiki-pm --skill llm-wiki-brief -a claude-code -gThis installs the skill with all its scripts, templates, and references. What you don't get without the full plugin: session hooks (auto-scaffold, prompt recall, wikilink validation, log rotation) and the bundled wiki-search. Core wiki operations (ingest, query, update, lint) work fine — you just need to create the wiki directory structure manually on first use.
git clone https://github.com/anh-chu/llm-wiki-pm ~/llm-wiki-pm
mkdir -p ~/.claude/skills
ln -s ~/llm-wiki-pm/skills/llm-wiki-pm ~/.claude/skills/llm-wiki-pmSame trade-offs as Option B. For hook automation and auto-scaffold, also install the plugin:
claude plugin install llm-wiki-pm@anh-chu-pluginsRestart Claude Code after enabling the plugin. You will be prompted for your wiki path and domain. The SessionStart hook creates the wiki directory on the first session start.
If you used Option B without installing the plugin, set WIKI_PATH before starting Claude Code:
echo 'export WIKI_PATH=$HOME/pm-wiki' >> ~/.bashrc && source ~/.bashrcFull setup, including mobile Obsidian sync, in GETTING_STARTED.md.
| llm-wiki-pm | kfchou/wiki-skills | lewislulu/llm-wiki-skill | lucasastorian/llmwiki | NotebookLM | |
|---|---|---|---|---|---|
| Shape | Single skill | 5 skills | Skill + plugin + server | Full web app | SaaS |
| Storage | Plain markdown | Plain markdown | Plain markdown | Supabase + S3 | Cloud |
| Search | Bundled BM25 + regex + tags (wiki-search) + backlinks | grep + index | grep + index | PGroonga | Proprietary |
| Update discipline | Diffs + supersession fields + auto-link rewrite | Diffs + source cite | Human-in-loop audit | None explicit | N/A |
| Privacy | Pre-ingest filter + private by default (shareable: allowlist) |
None | None | User-scoped | SaaS ToS |
| Transcript support | crystallize flow (decisions + actions) |
Generic ingest | Generic ingest | Generic ingest | Source-only |
| Install target | Claude Code, plus ChatGPT via wiki-mcp | Claude Code | OpenClaw / Codex | Self-host web | SaaS |
| Ops burden | None (local files) | None | Obsidian plugin + Node server | Supabase + S3 + OCR | Zero |
| Scales to 1000+ pages | Yes (wiki-search, bundled) | Degrades | Degrades | Yes | Yes |
| PM-tuned taxonomy | Yes (competitive, customer, strategy, roadmap, ai) | No | No | No | No |
For a deeper breakdown of which Karpathy and Rohit v2 ideas this implements, see the design notes below.
Retrieval quality was measured on a real working wiki (about 700 Markdown files, 11 MB, 325 of them raw sources) with 26 known-answer questions written from the raw sources: obscure facts, paraphrases, multi-document answers, facts that changed over time, and renamed concepts. Scoring is automatic (no LLM judge). Queries were short keyword or natural-language phrases written blind by a separate agent, the way an agent would search.
Hit@10 means a file holding the answer appears in the top 10 results. On this corpus the two plain-files engines came out ahead, at similar quality, with no database involved. Credit for that goes to the search projects themselves: @wirux/mcp-markdown-vault and obsidian-mcp, which now backs both the bundled wiki-search and wiki-mcp. This skill adds the structure and update rules around them, not the ranking.
Read these numbers with some care:
- 26 questions is a small set. One question is about 4 points, so gaps under roughly 10 points are noise.
- Questions about facts that changed over time or concepts that were renamed scored lowest for every backend tested (about half). Search alone does not fix those; they depend on the supersession and update rules in this skill.
- On a fresh copy of a wiki, mcp-markdown-vault's first full index stopped at about half the files with no error. That, its memory use, and its open write bugs are why 2.23 switched the bundled search to obsidian-mcp, which indexed every page in about 3 seconds.
Good fit if you:
- Work as a PM, analyst, researcher, or founder with lots of meetings and reports
- Want a local-first, markdown-based knowledge base that you own
- Use Claude Code as your primary agent, optionally with ChatGPT as a second client
- Are comfortable on a terminal (you'll occasionally run
lint.py, and ChatGPT access means running a small server behind a reverse proxy)
Not a fit if you:
- Want a zero-terminal SaaS → use NotebookLM
- Need team collaboration out of the box → use Notion or a shared Obsidian vault
- Don't use Claude Code → ChatGPT can read and write through wiki-mcp, but the skills, hooks, and sub-skill workflows are Claude Code only.
AGENTS.mdis the portable reference if you want to port to another agent
llm-wiki-pm/
├── AGENTS.md # Universal agent behavioral contract
├── CONTRIBUTING.md
├── .claude-plugin/
│ ├── plugin.json
│ └── marketplace.json
├── hooks/
│ └── hooks.json
├── scripts/
│ └── update-safe.sh
└── skills/
├── llm-wiki-pm/ # Core skill (required)
│ ├── SKILL.md
│ ├── hooks/ (session-start.sh, post-write.sh, session-stop.sh)
│ ├── references/
│ ├── scripts/ (lint.py, backlinks.py)
│ └── templates/ (SCHEMA.md, index.md, overview.md, log.md, persona.md,
│ MY-INTEGRATIONS.md)
├── llm-wiki-brief/ # Optional: daily/weekly briefs, tag digests
│ └── SKILL.md
├── llm-wiki-research/ # Optional: research sprints, competitive deep dives
│ └── SKILL.md
└── llm-wiki-maintain/ # Optional: daily-maintenance loop (sweep→ingest→brief→tidy)
└── SKILL.md
Two scenarios documented in detail in GETTING_STARTED.md:
- Human user with Claude Code: enable plugin, wiki auto-scaffolds on first session
- Application-orchestrated: programmatic wiki provisioning for platform deployments
Enable the plugin. When prompted, enter your wiki path and domain.
On the first session start, the SessionStart hook creates:
~/pm-wiki/ with SCHEMA.md, index.md, log.md, overview.md, and
the raw/entities/concepts/comparisons/queries/_archive subdirectories.
Ranked keyword (BM25), regex, tag and wikilink search over your wiki is
bundled as wiki-search, backed by
obsidian-mcp. The first session
downloads a pinned, checksum-verified binary (21 MB download, 64 MB
unpacked) to ~/.cache/llm-wiki-pm/; if obsidian-mcp 3.x is already on
your PATH it is used instead. Each Claude Code session runs its own copy,
which builds the index in memory in seconds at startup. obsidian-mcp keeps
its data folder in ~/.local/share/obsidian-mcp/ but still creates a small
.obsidian-mcp/ folder in the wiki (add it to .gitignore). Needs
python3 (already required by the hooks), which trims search results to a
few snippets per file. _archive/ snapshots are left out of search;
note_read still opens them by path.
Open ~/pm-wiki/SCHEMA.md. Adjust:
- Domain statement (scope)
- Tag taxonomy (add/remove tags for your specific accounts, competitors, themes)
- Page thresholds (tune later after a few ingests)
Option A, user-level (available in every project):
mkdir -p ~/.claude/skills
ln -s /home/sil/llm-wiki-pm/skills/llm-wiki-pm ~/.claude/skills/llm-wiki-pmOption B, project-level (per-repo):
mkdir -p .claude/skills
ln -s /home/sil/llm-wiki-pm/skills/llm-wiki-pm .claude/skills/llm-wiki-pmRestart Claude Code. Verify with /skills, llm-wiki-pm should appear.
The skill auto-activates on ingest/query/update/lint phrasing (see SKILL.md
"When This Skill Activates").
In Claude Code:
"Ingest this Gartner Magic Quadrant report: "
Claude reads SKILL.md, orients on SCHEMA + index + log + overview, surfaces takeaways, creates/updates pages, logs.
python3 /home/sil/llm-wiki-pm/skills/llm-wiki-pm/scripts/lint.py ~/pm-wiki
# opens queries/lint-YYYY-MM-DD.mdSee references/obsidian-sync.md for obsidian-headless + systemd setup.
Weekly competitive digest: ingest 3-5 analyst links in one session. Batch updates, one log entry, refresh overview.md.
Pre-meeting prep: query "what do we know about ?" → Claude
reads entities/<customer>.md + recent log → offer to file post-meeting update.
Monthly 1:1 follow-up: ingest transcript → extract decisions/themes → update relevant concept pages → link from person page.
Quarterly review: lint, triage, rotate log, refresh overview, prune tag taxonomy.
Wiki = long-term curated facts + sources you review, cite, and share with colleagues. Not a replacement for session memory or notes apps. If you pair with a memory tool (Claude's native memory, mem0, Hindsight), keep them non-overlapping: wiki holds facts + sources, memory holds persona + session state.
Hook scripts and plugin manifest validation:
python3 -m venv .venv && .venv/bin/pip install pytest -q
.venv/bin/python -m pytest tests/ -v43 tests covering scaffold, wikilink validation, log rotation, stdin parsing, and plugin manifest compliance. All tests create isolated temp wikis and feed the real Claude Code hook JSON schema to the scripts.
MIT.
Built on prior art from:
- Andrej Karpathy, LLM Wiki , the original pattern: stop re-deriving, start compiling. Three-layer architecture, Memex lineage, and the insight that LLMs are the first librarians who don't get bored of bookkeeping.
- Rohit G, LLM Wiki v2 , lifecycle concepts (supersession, privacy, crystallization, self-healing lint). We cherry-picked the four highest-ROI v2 ideas for PM work.
- kfchou/wiki-skills: update
discipline with diffs and stale-claim sweep, tiered lint reports, evolving
overview.mdsynthesis. - lewislulu/llm-wiki-skill , audit/feedback loop design (inspiration for future team-mode support).
- obsidian-mcp: BM25, regex and wikilink search over markdown vaults (bundled as wiki-search, and the backend of wiki-mcp).
- @wirux/mcp-markdown-vault: semantic + TF-IDF search, bundled as wiki-search up to 2.22.
How this skill maps to Karpathy's original gist and Rohit's v2 extensions:
- Three-layer architecture (raw sources, agent-owned wiki, schema)
- LLM owns the wiki; human curates sources
- Ingest / query / lint operations
index.mdcontent catalog +log.mdchronological record- File good answers back as pages (
queries/dir) - Obsidian compatibility (Graph, Dataview, frontmatter)
- Schema as key configuration, co-evolved
- Ingests touch 10-15 pages routinely
- Bundled wiki-search via
obsidian-mcp - Multi-format outputs: Marp, matplotlib, CSV, Mermaid, Canvas
Implemented:
- Explicit supersession with
supersedes:/superseded_by:fields + auto-redirect - Privacy filter (pre-ingest checklist + private-by-default
shareable:allowlist) - Self-healing lint (
--auto-fixfor safe repairs) - Crystallization (transcript → decision digest)
- Schema as the real product
- Contradiction handling with frontmatter flag
- Backlink tracing (
scripts/backlinks.pyfor structural refs)
Intentionally skipped for solo PM use (overkill):
- Confidence decay curves
- Consolidation tiers (working/episodic/semantic/procedural memory)
- Typed knowledge graph with relationship types
- Multi-agent mesh sync
- Quality scoring pipeline