← Docs hub

Slash commands reference

Every /wiki-* command llmwiki install-agent-kit ships — what it does, what it runs under the hood, and a realistic invocation example. Use these inside Claude Code. The command files live in the installable package and land in an agent directory via llmwiki install-agent-kit --dest PATH.

Maintainer and AWOS delivery commands (/release, /fix-bug, /implement-feature) are not part of the vault pipeline and are not installed by the agent kit — they are described in ../maintainers/README.md.

All 12 commands in the vault pipeline, in the order you meet them:

Command What it does
/wiki-init Scaffold an empty vault (raw/, wiki/, site/)
/wiki-sync Convert new agent sessions into raw/
/wiki-ingest Ingest one file or folder into raw/
/wiki-synth Synthesize pending raw into wiki/sources/, then harvest candidates
/wiki-candidates Triage pending candidate stubs
/wiki-query Answer a free-form question from the wiki
/wiki-update Edit one wiki page in place
/wiki-lint Check wiki quality — orphans, broken links, stale pages
/wiki-graph Build the knowledge graph from [[wikilinks]]
/wiki-reflect Higher-order reflection pass over the whole wiki
/wiki-build Regenerate the static HTML site
/wiki-all Run the whole pipeline end-to-end

Decision tree: which tool runs when?

CLI vs slash

You want to… Use
…run a check in CI, a cron job, or a shell script CLI (python3 -m llmwiki …)
…chain commands with && / pipe to jq CLI
…have the model read output + take follow-up actions slash (inside Claude Code / Codex)
…answer a free-form question ("what did I decide about X?") slash (/wiki-query)
…do one-shot builds, graph generation either — slashes wrap the CLI

Rule of thumb: if the output is for you to read + act on manually, use the CLI. If the output should feed back into an LLM turn, use the slash — the model sees the full stdout and can chain into the next step.

Lint (wiki quality)

Structural and content quality for the wiki is llmwiki lint / /wiki-lint — there is no separate eval subcommand.

Command Checks Severity model When to run
llmwiki lint / /wiki-lint Frontmatter completeness, [[wikilink]] integrity, orphans, duplicate titles, stale pages, tag-topic convention, stale references, findability (page_findability, title_ambiguity, search_consistency), and the rest of the registered rules Rules with error / warning / info severities; --fail-on-errors exits non-zero only on errors After every /wiki-sync or /wiki-build, and in CI

Reach for lint when a page or the corpus looks wrong: orphans, broken [[wikilinks]], missing frontmatter, stale summaries. Use --fail-on-errors (or the automation lint-fail policy) when a non-zero exit should block a pipeline.


Wiki pipeline

/wiki-init

What: scaffolds an empty llmwiki — creates raw/, wiki/, site/ and seeds wiki/index.md, wiki/log.md, wiki/overview.md, plus the nine navigation files (CRITICAL_FACTS.md, MEMORY.md, SOUL.md, hints.md, hot.md, dashboard.md).

Wraps: python3 -m llmwiki init.

When to use: first time in a fresh repo, or after deleting wiki/ to start over.

Example:

/wiki-init

Claude Code will respond by running init and surfacing which files were seeded.


/wiki-sync

What: convert new Claude Code (+ Codex + Cursor + etc.) .jsonl sessions into markdown under raw/sessions/, then ingest into wiki/.

Wraps: python3 -m llmwiki sync.

Arguments Claude may pass through: --dry-run, --since, --project, --force, --vault. Say any of them in natural language — "sync but only sessions from this week" becomes --since $(date -v-7d +%Y-%m-%d). Durable lookback in config (filters.since / adapters.*.since) applies on bare /wiki-sync when --since is omitted — see configuration-reference.md.

When to use: at the end of each coding block. Also the only command that triggers auto-ingest of new pages into wiki/.

Example:

/wiki-sync
/wiki-sync only my llm-wiki project
/wiki-sync but don't auto-build afterwards
/wiki-sync into my Obsidian vault at ~/Documents/Obsidian Vault

Expected output (narrated):

==> claude_code: 3 new sessions since last sync
✓ wrote 3 pages under raw/sessions/
✓ ingested into wiki/sources/
✓ auto-build: site/ rebuilt (690 HTML files)

/wiki-ingest <path>

What: ingest one source document or folder into wiki/, or enrich / discuss pending candidates during review. Reads the file, creates / updates the matching wiki/sources/<slug>.md, and may propose entity/concept candidates. Trusted hubs still require review (/wiki-candidates or llmwiki candidates promote|merge|discard) — ingest is not an auto-promote escape hatch.

Wraps: the Ingest Workflow in CLAUDE.md (no single CLI — it's a slash-command-driven workflow that the model orchestrates).

When to use: you dropped a source file manually (a PDF, a Jira ticket export, a meeting transcript), or Home / Analytics show a To review backlog and you want agent-led discussion over candidates. For bulk stub generation from already-synthesized sources, prefer llmwiki synth --candidates-only.

Examples:

/wiki-ingest raw/sources/2026-04-17-incident.md
/wiki-ingest raw/jira/
/wiki-ingest ~/Downloads/meeting-transcript.vtt

/wiki-synth

What: synthesize pending raw sessions/docs into wiki/sources/, then harvest entity/concept candidates into wiki/candidates/ (default). Use --sources-only for the legacy sources-only path. Sources are two LLM jobs per run (known-names prepare + one ask per queued file); harvest is offline. Ctrl+C harvests from written pages (or prints synth --candidates-only after --sources-only) and exits 130. Do not run a separate consolidate-topics step — known-names prepare is part of synth.

Wraps: python3 -m llmwiki synth.

Example:

/wiki-synth
/wiki-synth with a cost estimate
/wiki-synth force a re-run of every source
/wiki-synth sources only

/wiki-candidates

What: triage pending candidates — promote, flip-promote, merge, discard, or batch apply --actions.

Wraps: python3 -m llmwiki candidates list + follow-ups (apply --actions for batches). Same intents site/candidates.html lists, whose copyable batch feeds the same command.

When to use: Home Candidates / Analytics Candidates to review is non-zero, /wiki-lint reported stale_candidates, or you just ran llmwiki synth / synth --candidates-only.

Promote fills an empty ## Key Facts offline from source fact: bullets (and harvest stubs); Dummy / no backend is fine (#147). Prefer the CLI action for the common case. Opt-in llmwiki candidates rewrite-key-facts --slug <Name> (or --all) still needs an LLM for trusted pages with regex-era Key Facts or pasted harvest-stub ## Candidate merge blocks. Prefer flip-promote over hand-moving stubs between candidates/entities and candidates/concepts.

Example:

/wiki-candidates

Claude will walk the queue one at a time and offer actions per candidate.


/wiki-query <question>

What: answer a question from the wiki. Reads wiki/index.md + wiki/overview.md + any cache_tier: L1 pages, then walks relevant source / entity / concept pages and synthesises an answer with inline [[wikilinks]] back to the originals.

Wraps: the Query Workflow in CLAUDE.md.

When to use: "have I solved this before?" / "when did I add X?" / "which sessions touched Y?".

Examples:

/wiki-query when did I add the lint rules?
/wiki-query which agent did I use for refactoring the cache-tier module?
/wiki-query summarize every session about Obsidian integration

Save prompt: if the answer runs 3+ paragraphs, Claude will offer to save it under wiki/syntheses/<slug>.md.


/wiki-update <page>

What: surgically edit one wiki page without re-ingesting. Useful for fixing broken wikilinks, updating stale frontmatter, adding a missing ## Connections line.

When to use: lint flagged something, you know the fix, you don't want to re-run sync.

Example:

/wiki-update wiki/entities/RAG.md add a Connections section linking to Karpathy and llm-wiki

/wiki-lint

What: run every registered lint rule (all deterministic — no LLM). The live number is printed by llmwiki lint --help.

Wraps: python3 -m llmwiki lint.

Rules, in order:

  1. frontmatter_completeness
  2. frontmatter_validity
  3. link_integrity
  4. orphan_detection
  5. content_freshness
  6. duplicate_detection
  7. index_sync
  8. contradiction_detection — non-filler ## Contradictions sections
  9. claim_verification — entity/concept claims without sources
  10. summary_accuracy — empty summary: frontmatter
  11. stale_candidates
  12. tags_topics_convention (G-16 · #302)
  13. stale_reference_detection (G-17 · #303)
  14. frontmatter_count_consistency
  15. tools_consistency
  16. stub_source_pages
  17. provenance_integrity (#122)
  18. page_findability (#197) — titled page not returned for its own title (findability key is title, not slug; prefer [[slug|Title]])
  19. title_ambiguity (#197) — titled page not ranked first for its own title
  20. search_consistency (#197) — search vs literal scan; survival share is informational, not a defect

See cli.md lint for severities and findability detail.

Example:

/wiki-lint
/wiki-lint just the link_integrity rule

/wiki-graph

What: build the knowledge graph. Nodes = wiki pages, edges = [[wikilinks]]. Emits graph/graph.json + graph/graph.html.

Wraps: python3 -m llmwiki graph.

Example:

/wiki-graph

Then open site/graph.html (auto-copied from graph/graph.html during build) in a browser.


/wiki-reflect

What: higher-order self-reflection pass over the whole wiki. Looks for gaps, patterns, duplicated-topic clusters, areas where a synthesis page would help.

No CLI wrapper — it's a model-orchestrated workflow that reads the index + overview + sample of pages and outputs suggestions.

Example:

/wiki-reflect

Use sparingly; it's the most token-heavy command.


/wiki-build

What: regenerate the static HTML site.

Wraps: python3 -m llmwiki build.

When to use: after manual edits to wiki/, or when you want to see a fresh site without running the full sync pipeline.

Example:

/wiki-build
/wiki-build to ~/public_html
/wiki-build in tree search mode

/wiki-all

What: run the full pipeline end-to-end — sync → synth → build → graph → lint. Every stage runs unless you opt out of it. AI-consumable exports (llms.txt, sitemap.xml, etc.) are written by build, not a separate step.

Wraps: python3 -m llmwiki all.

When to use: after /wiki-sync, when you want a CI-ready site in one shot instead of chaining /wiki-build + /wiki-graph + /wiki-lint yourself.

Example:

/wiki-all
/wiki-all --no-synth
/wiki-all --graph-engine builtin
/wiki-all --skip-graph --strict

Pass --strict to turn any lint warning into a non-zero exit, which is exactly what CI wants. Pass --skip-graph or --graph-engine builtin when the optional Graphify backend is not installed. Pass --no-sync or --no-synth to leave session conversion or synthesis out of the run — --no-synth is the one that keeps the run away from your AI provider.


Agent-kit skills

The kit also ships four skills — the model invokes these on its own when what you ask for matches their description, so they cover the same work without you remembering a slash command. They land under skills/ in the same --dest as the commands.

Skill What it does Typical trigger
llmwiki-sync Converts new agent sessions into raw/sessions/ (llmwiki sync), then ingests what arrived "sync the wiki", "catch me up", a question that needs recent sessions
llmwiki-ingest Ingests one document or folder: add (CLI or MCP wiki_add) → synth → candidate review → build. Documents never become hand-written pages "ingest this", "add this to the wiki", /wiki-ingest
llmwiki-query Answers a question from the wiki, with [[wikilink]] citations, and offers to save a substantial answer under wiki/syntheses/ "what did I decide about X?", /wiki-query
llmwiki-all Runs the whole pipeline end-to-end — sync → synth → build → graph → lint — and reports each stage "run everything", "full pipeline", /wiki-all

The skill is named llmwiki-all; the slash command it wraps is /wiki-all. Upgrading from a release that shipped the skill as wiki-all prunes the old copy when you re-run llmwiki install-agent-kit --dest PATH (the prune is by content, so a copy you edited is left alone and reported as kept).


How the slash commands get installed

llmwiki install-agent-kit --dest PATH copies the packaged wiki-*.md command files and the skills/ folder into an agent directory — ~/.claude (Claude Code), ~/.cursor (Cursor), ~/.codex (Codex CLI) for every project on the machine, or .claude for just the project you are working in. The agent picks them up from there with no further setup.

For an agent with a different layout, point --dest at (or copy the installed files into) its own command and skill directories — the file format is portable. The destination table and the pruning rules are in cli.md.


Extending

To add a new slash command:

  1. Create llmwiki/agent_kit/commands/wiki-<name>.md with a one-line docstring on line 1 (that's the summary Claude Code surfaces).
  2. Describe the workflow in prose. Reference existing CLI commands rather than embedding shell in the body.
  3. Run /wiki-lint — the docs/reference/ guardrail test (see tests/test_docs_structure.py) will pick up the new command.
  4. Document it here — the CI guard requires every llmwiki/agent_kit/commands/*.md file to have a matching ### entry, and the count line above to match how many there are. A maintainer-only command goes in .claude/commands/ and is described in ../maintainers/README.md instead.

Keyboard shortcuts

⌘K / Ctrl+KOpen command palette
/Focus search
g hGo to home
g pGo to projects
g sGo to sessions
j / kNext / prev row (tables)
?Show this help
EscClose dialogs

Structured queries

Mix key:value filters with free text in the palette:

type:sessionOnly session pages
project:llm-wikiFilter by project name (substring)
model:claudeFilter by model name (substring)
date:>2026-03-01Sessions after a date
date:<2026-04-01Sessions before a date
tags:rustPages mentioning a tag/topic
sort:dateSort results by date (newest first)

Example: type:session project:llm-wiki date:>2026-04 sort:date