Help us improve
Share bugs, ideas, or general feedback.
Share bugs, ideas, or general feedback.
Share bugs, ideas, or general feedback.
By genkovich
LLM Wiki workflows — ingest, query, lint, inbox. Operates on any wiki whose CLAUDE.md has frontmatter type: schema, scope: wiki.
npx claudepluginhub genkovich/llm-wiki --plugin wikiUse when the user wants to drop a source into the LLM Wiki inbox WITHOUT parsing it yet. Trigger phrases: "wiki:inbox <url|path>", "inbox this", "drop in inbox", "save to wiki later", "wiki: queue <url>", or when the user pastes a URL/file and says "for later". Stages the input in raw/_inbox/ so the wiki:parse skill (or the cron-ingest launchd job) can process it asynchronously. This skill does NOT call defuddle/yt-dlp/whisper — it only stages. Heavy parsing happens later.
Use when the user invokes LLM Wiki lint workflow — explicitly via "wiki:lint", "wiki: lint", "перевір wiki", "audit the wiki", or similar consistency-check requests. Scans entities/, concepts/, and comparisons/ for contradictions, stale pages, orphans, missing cross-references, and data gaps. Writes findings into a fresh questions/lint-YYYY-MM-DD.md and appends a log entry. Recognizes the wiki by a CLAUDE.md with `type: schema, scope: wiki` and follows the lint protocol from there.
Use when the user invokes LLM Wiki parse workflow — explicitly via "wiki:parse", "parse this", "parse this folder", "wiki: parse <path|url>", drops a YouTube link, or when a new file is dropped into a wiki's raw/_inbox/. Handles four input kinds: single file, folder (batch), generic URL (defuddle clip), or YouTube URL (yt-dlp transcript + optional whisper). Processes each source into the wiki: creates sources/<slug>.md summary, updates 10-15 entities/concepts with citations, appends a log entry. Recognizes the target wiki by a CLAUDE.md with `type: schema, scope: wiki` frontmatter. Reads that CLAUDE.md for the authoritative ingest protocol before acting.
Use when the user invokes LLM Wiki query workflow — explicitly via "wiki:query", "wiki: <question>", "що каже wiki про X", "знайди в wiki", "ask the wiki", or similar retrieve-and-synthesize requests against an LLM Wiki. Performs MCP qmd search across the wiki collection, reads the top-N results, and returns a synthesis with inline wikilink citations. Optionally proposes a filer-back into comparisons/ or questions/. Recognizes the wiki by a CLAUDE.md with `type: schema, scope: wiki` frontmatter and follows the query protocol from that file.
Share bugs, ideas, or general feedback.
Own this plugin?
Verify ownership to unlock analytics, metadata editing, and a verified badge.
Sign in to claimOwn this plugin?
Verify ownership to unlock analytics, metadata editing, and a verified badge.
Sign in to claimBased on adoption, maintenance, documentation, and repository signals. Not a security audit or endorsement.
Comprehensive C4 architecture documentation workflow with bottom-up code analysis, component synthesis, container mapping, and context diagram generation
Harness-native ECC plugin for engineering teams - 64 agents, 262 skills, 84 legacy command shims, reusable hooks, rules, MCP conventions, and operator workflows for Claude Code plus adjacent agent harnesses
Tools to maintain and improve CLAUDE.md files - audit quality, capture session learnings, and keep project memory current.
Upstash Context7 MCP server for up-to-date documentation lookup. Pull version-specific documentation and code examples directly from source repositories into your LLM context.
Manus-style persistent markdown files for planning, progress tracking, and knowledge storage. Works with Claude Code, Kiro, Clawd CLI, Gemini CLI, Cursor, Continue, Hermes, and 17+ AI coding assistants. Now with Arabic, German, Spanish, and Chinese (Simplified & Traditional) support.
v9.44.0 — Patch release for cursor-agent smoke checks in untrusted workspaces. Run /octo:setup.
Spec-Driven Development for Claude Code: 16 atomic, stack-agnostic skills + 9 subagents that carry a feature from idea → spec → design → data-model → API → tasks → tests → implement → review → ship, plus a TDD implementation engine and a living roadmap above the per-feature flow. Start with `survey` (maps an existing codebase or bootstraps an empty one), walk the gated backbone — each stage reads the previous artifact and hard-refuses if missing — and keep `docs/roadmap.md` (Now/Next/Later, outcome-altitude, never dates) in sync as `specify` promotes a feature to Now and `ship` moves it to Shipped. The Q&A skills are depth-tunable (easy/medium/hard — decide-for-me vs. interrogate-me-with-trade-offs); the ideation analyses run as named subagents (researcher/strategist/analyst/devil's-advocate) gated by that dial; diagrams are presented in prose and written to file (never raw Mermaid dumped to the terminal); and every acceptance criterion is covered end-to-end. design is surface-aware: it declares the feature's target surfaces at the architecture stage (backend-service / web-frontend / mobile-app / desktop-app / cli / worker / library-sdk, each a C4 container) and downstream stages gate their output by that choice — the `ui` task layer, UI-driven sequence flows, and the component / visual-regression / e2e-through-UI frontend test tiers; `ui` work reuses the repo's existing design system (components / tokens / styling) from the architecture map, never reinvents. Every stage ends with a copy-ready handoff block — what changed, the files to review, then `/clear` + the next `/sdd:…` command.
Enforce team PR review conventions: changelog entry, no console.log, test coverage, no formatting churn, security checklist.
Pre-deploy checklist enforcer: env diff, migrations dry-run, smoke endpoints, rollback plan, oncall heads-up.
Pre-deploy checklist enforcer: env diff, migrations dry-run, smoke endpoints, rollback plan, oncall heads-up.
Enforce team PR review conventions: changelog entry, no console.log, test coverage, no formatting churn, security checklist.
Personal LLM-curated wiki: a structured markdown KB that an LLM grows incrementally (ingest → sources → entities/concepts) instead of re-deriving knowledge each time through RAG. Local search via qmd + MCP integration with Claude Code.
This repository is both a Claude Code marketplace (so you can install the wiki plugin with one command) and a template scaffold (init.sh materializes a fresh wiki in any folder).
# 1. Materialize a wiki skeleton inside your project (no clone — fetches template tarball)
curl -fsSL https://raw.githubusercontent.com/genkovich/llm-wiki/main/init.sh | bash -s ./wiki
# 2. Add this repo as a Claude Code marketplace and install the plugin
# Inside Claude Code:
/plugin marketplace add genkovich/llm-wiki
/plugin install wiki@llm-wiki
# 3. Install qmd (search engine: on-device BM25 + vector + LLM rerank)
npm install -g @tobilu/qmd
qmd collection add ./wiki --name my-wiki
qmd context add qmd://my-wiki "<1-2 sentences about the domain — affects reranking>"
qmd embed # first run downloads ~330MB of embedder GGUF
In Claude Code, install the qmd plugin so that query/get/multi_get are available as native MCP tools:
/plugin marketplace add tobi/qmd
/plugin install qmd@qmd
/reload-plugins
After that the skills are reachable under the wiki: namespace:
| Skill | Triggers on |
|---|---|
wiki:inbox | inbox this, wiki: inbox <url>, drop a URL for later parsing |
wiki:parse | parse this, wiki: parse <path>, a new file in raw/_inbox/ |
wiki:query | wiki: <question>, що каже wiki про X, знайди в wiki |
wiki:lint | wiki: lint, перевір wiki, audit the wiki |
Each skill locates the target wiki by type: schema, scope: wiki in the CLAUDE.md frontmatter and runs the workflow from that wiki's CLAUDE.md.
Auto-parse raw/_inbox/ every 10 minutes — through launchd on macOS or systemd user units on Linux.
mkdir -p ~/.config
WIKI_PLUGIN_BASE="$HOME/.claude/plugins/cache/llm-wiki/wiki"
WIKI_PLUGIN="$WIKI_PLUGIN_BASE/$(ls -1t "$WIKI_PLUGIN_BASE" | head -1)"
cp "$WIKI_PLUGIN/scripts/wiki-parse.env.example" ~/.config/wiki-parse.env
# edit ~/.config/wiki-parse.env — set WIKI=/path/to/your/wiki
sed "s|__PLUGIN_ROOT__|$WIKI_PLUGIN|g" "$WIKI_PLUGIN/launchd/com.wiki-parse.plist.example" \
> ~/Library/LaunchAgents/com.wiki-parse.plist
launchctl load ~/Library/LaunchAgents/com.wiki-parse.plist
Logs: $WIKI/.cron-parse.log. Stop: launchctl unload ~/Library/LaunchAgents/com.wiki-parse.plist.
mkdir -p ~/.config/systemd/user
for f in wiki-parse.service wiki-parse.timer; do
sed "s|__PLUGIN_ROOT__|$WIKI_PLUGIN|g" "$WIKI_PLUGIN/systemd/$f.example" \
> ~/.config/systemd/user/$f
done
systemctl --user daemon-reload
systemctl --user enable --now wiki-parse.timer
Logs: ~/.local/state/wiki-parse.log and $WIKI/.cron-parse.log. Stop: systemctl --user disable --now wiki-parse.timer.
Working on the plugin itself? Add the local checkout as a marketplace and reinstall after every change:
/plugin marketplace add /path/to/llm-wiki
/plugin install wiki@llm-wiki
/reload-plugins
After init.sh runs, open and edit the placeholders in the new wiki:
wiki/CLAUDE.md — replace <DOMAIN> and <DOMAIN_DESCRIPTION> with your domain. This is the schema the LLM reads on every ingest/query/lintwiki/overview.md — set the domain title in the headingwiki/log.md — set today's date in the bootstrap entryThe four skills above map to these actions — authoritative protocols live in the generated wiki/CLAUDE.md, which the LLM reads on every invocation:
wiki:inbox — queue helper: drop a URL or file into raw/_inbox/ for later parsing (no other writes)wiki:parse — ingest pipeline: raw source → sources/<slug>.md → updates 10-15 entities//concepts/ with citations → append log.mdwiki:query — MCP qmd search → synthesis with [[wikilink]] citations → optional filer-back into comparisons//questions/wiki:lint — find contradictions, orphans, stale pages, data gaps → write into questions/lint-YYYY-MM-DD.md