← 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:

Decision tree: which tool runs when?

CLI vs slash

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.

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 (16 at last count — all structural / deterministic). 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

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.

How the slash commands get installed

llmwiki install-agent-kit --dest PATH copies the packaged wiki-*.md command files into an agent directory — --dest .claude for the project you are working in, or a user-level agent directory. Claude Code picks them up from there with no further setup.

For Codex CLI / Cursor / Gemini CLI / other agents, point --dest at (or copy the installed wiki-*.md files into) the corresponding skill directory for that agent (typically .codex/skills/ or .agents/skills/) — the file format is portable across agents.

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.

Related

• CLI reference — the underlying python3 -m llmwiki … surface.
• UI reference — every screen on the compiled site, with what's reachable from where.
• Tutorial 03 — Use with Claude Code — the minimum daily loop built on these commands.