Skills

Echo Wiki uses Agent Skills to manage the wiki pipeline. Skills are stored in .claude/skills/ and work with any compatible agent.

After onboarding, wiki operations run a structure check (Step 0) before starting. If any required wiki paths are missing, the skill recreates them automatically. See _meta/prompts/structure-check.md for details.

Wiki operations append an entry to wiki/_log.md after completing — a chronological, parseable record of every operation. The log is auto-created on first use and preserved across rebuilds.

/onboard

Initialize a fresh or existing Git repository without relocating its artifacts.

/onboard /absolute/path/to/company-kb

The skill inspects the repository and loads the relevant fresh/existing guide. A script previews and installs the standard runtime; the agent handles configuration, instruction/hook integration and one selected source. Conflicts produce actionable next steps. Reruns preserve existing instances; no automatic upgrade, commit or push occurs.

For first-time use before the skill is installed, follow the explicit skill-file examples in Getting Started.

Keeping Content Fresh

Echo Wiki is command-driven, not a background synchronizer. /ingest updates the KB pipeline for a new source, but context packs, durable answers, knowledge gaps, and the maintenance queue are derived artifacts that refresh only when their respective skills run. See Keeping Content Fresh for the change-to-next-step guide, including the deliberate workflow for replacing a source.

/ingest

Fetch and clean source content into raw/.

/ingest <url>           # Ingest a web URL
/ingest <file-path>     # Ingest a local file (md, pdf)

What it does:

1. Detects source type from URL pattern (blog, substack, github, paper, tweet); company-written Markdown and text files are filed as internal
2. Fetches content via Tavily or Firecrawl
3. Preserves source headings or adds ## Content when the cleaned body has no heading, ensuring every new raw source has a stable evidence anchor
4. Downloads images locally
5. Writes clean markdown with frontmatter to raw/
6. Validates the raw file with hooks/validate.sh
7. Appends entry to wiki/_log.md
8. Automatically triggers /compile

Source type detection:

URL Pattern	Type	Directory
*.substack.com/*	substack	raw/substacks/
github.com/*	github	raw/github/
twitter.com/*, x.com/*	tweet	raw/people/
arxiv.org/*, *.pdf	paper	raw/papers/
Other URLs	blog	raw/blogs/
Local .md / .txt authored by your team	internal	raw/internal/
Podcasts / videos	—	User must specify type
Team-authored PDF or URL	internal (override)	raw/internal/

/compile

Compile raw sources into structured wiki articles.

/compile raw/<category>/article.md   # Compile a specific source
/compile all                         # Recompile entire wiki

What it does:

1. Reads raw source(s)
2. Creates source-summary in wiki/sources/
3. Extracts entities based on configured entity_types (default: concepts, people, tools)
4. Creates new articles or merges into existing ones (never overwrites)
5. Adds [[wikilinks]] between related articles
6. Runs hooks/reindex.sh to regenerate _index.md and _backlinks.md deterministically (includes workspace content)
7. Validates written articles with hooks/validate.sh, fixing violations before finishing
8. Appends entry to wiki/_log.md

KB entity types (configurable in _meta/wiki.config.yaml):

Type	Directory	Examples
Concepts	wiki/concepts/	Ideas, theories, patterns
People	wiki/people/	Researchers, authors, key figures
Tools	wiki/tools/	Software, platforms, frameworks
Sources	wiki/sources/	Summary of each raw source

These are the defaults. Custom wikis can define different entity types — see Configuration.

/compile stops without writing KB content when a legacy raw source has no visible heading. Because skills treat raw/ as append-only, migrate that source deliberately by adding ## Content before its body, then retry compilation.

/rebuild

Stage and validate a complete KB replacement from all remaining raw sources.

/rebuild

Use this after manually deleting one or more raw source files. The /compile skill only appends and merges — it cannot remove content from deleted sources. /rebuild starts fresh and recompiles only from sources that still exist.

What it does:

1. Collects all remaining raw sources (raw/**/*.md)
2. If no sources found, aborts safely — KB directories are not wiped
3. Validates every collected raw source and aborts without modifying wiki/ if any source cannot be replayed
4. Acquires an exclusive rebuild lock and creates a checksummed, isolated staging wiki with empty configured KB type directories and byte-for-byte copies of preserved content
5. Replays each source chronologically (ingested date, oldest first) using the compile workflow; any replay failure aborts and discards staging
6. Runs hooks/reindex.sh and hooks/validate.sh --all against staging, including preserved workspace dependencies
7. If staging fails, leaves the live wiki unchanged and reports obsolete context, queue, gap, or workspace dependencies for explicit remediation
8. After a final snapshot comparison, journal-swaps the complete wiki/ directory and retains the old version until post-install validation succeeds
9. Never modifies wiki/workspaces/, wiki/.obsidian/, or existing wiki/_log.md content, then appends the rebuild summary

Removing a source from the wiki:

# 1. Delete the raw source file
rm raw/substacks/outdated-article.md

# 2. Rebuild to reconcile
/rebuild

After rebuild, all articles unique to the deleted source are gone, and multi-source articles are rewritten without the deleted source's content. Workspace content is untouched.

If a preserved workspace artifact still cites the deleted source or links to an article that is no longer produced, staging validation aborts before live files change. Regenerate that derived artifact with /context, /query, or /maintain, or explicitly remove it if it is obsolete, then retry.

The rebuild transaction helper detects concurrent changes before commit. If the process is interrupted during the directory swap, the next /rebuild restores the previous complete wiki from its recovery marker before starting again.

TIP

raw/ is append-only during normal operations (/ingest and /compile never modify existing raw files). Only delete raw files as a deliberate manual action before running /rebuild.

/index

Rescan wiki/ and update _index.md and _backlinks.md.

/index

Use this after manually creating or modifying workspace content (notes, drafts, etc.) to update the master index.

What it does:

1. Runs hooks/reindex.sh, which scans all .md files in wiki/ (KB articles + workspace content)
2. _index.md is regenerated with entries grouped by type and workspace
3. _backlinks.md is regenerated with cross-zone references and explicit _No inbound links._ orphan markers
4. Appends entry to wiki/_log.md

This is a non-destructive, fully deterministic operation — no LLM writes the two index files, so they cannot drift or silently drop entries.

/lint

Run semantic validation checks on the wiki.

/lint                    # Lint entire wiki
/lint all                # Lint entire wiki (explicit)
/lint wiki/concepts/     # Lint specific directory
/lint --domain llm       # Lint articles tagged with a specific domain

Produces a report at output/reports/lint-<date>.md and appends a summary to wiki/_log.md. Runs 7 checks:

1. Deterministic validation — runs hooks/validate.sh --all (full schema, enums, dates, tags vs domains, source paths, filenames, wikilinks); violations become Critical Issues
2. Orphaned articles — collected from _backlinks.md's explicit _No inbound links._ markers
3. Contradictory claims — conflicting facts across related KB articles
4. Stale content — past decay rate threshold (KB articles only)
5. Missing concepts — topics mentioned in 3+ articles without their own article
6. Duplicate detection — same entity under different names
7. Source fidelity (sampled) — re-reads cited raw sources for a sample of articles and flags claims not traceable to any cited source

/query

Answer questions from the wiki, with citations — and keep the good answers.

/query <question>

What it does:

1. Navigates progressively: _index.md → relevant articles (raw sources only as a last resort)
2. Synthesizes an answer citing every article used with [[wikilinks]]
3. If the answer is durable (multi-article synthesis, likely to recur), resolves a trusted actor workspace and files it under answers/ with collision-safe naming, evidence validation, and reindexing
4. When evidence is insufficient or low-confidence, creates or updates a shared gap note in wiki/workspaces/knowledge-maintenance/gaps/
5. Appends entry to wiki/_log.md

Query answers never touch KB type directories or raw/: the KB stays a pure projection of raw/, and /rebuild can never destroy filed answers.

An unanswered engineering question becomes a durable gap note rather than a guessed answer. The note records the question, sources searched, missing evidence, and a suggested next primary source to ingest. Repeated questions update the same note so maintenance can prioritize what real users and agents need next.

For example, if “How are authentication tokens rotated?” is unsupported, /query creates one gap note for that question and recommends ingesting the primary authentication RFC, ADR, code document, or tracking issue. Asking again updates the same note instead of creating a duplicate.

/context

Create or refresh concise, evidence-backed working context for a product area or component.

/context authentication

What it does:

1. Loads the index, relevant articles, and cited raw sources progressively
2. Writes a rebuild-safe context pack to wiki/workspaces/knowledge-maintenance/context/, preserving distinct product areas when their slugs collide
3. Summarizes purpose, current architecture, constraints, decisions, open questions, and recommended next reading
4. Adds Evidence: raw/<path>.md#<exact heading> after factual paragraphs
5. Validates and reindexes the pack

Context packs are a focused Markdown view for coding work, not a separate graph or a replacement for source-backed KB articles.

/maintain

Refresh generated indexes and create a safe, prioritized maintenance queue.

/maintain

What it does:

1. Runs reindex.sh, deterministic validation, and the existing /lint all workflow
2. Reads shared knowledge gaps from wiki/workspaces/knowledge-maintenance/gaps/
3. Writes wiki/workspaces/knowledge-maintenance/maintenance-queue.md, ordered by broken evidence, contradictions/staleness, recurring gaps, then orphan/duplicate candidates
4. Validates and reindexes the generated queue

/maintain is safe autopilot. It refreshes generated indexes, writes the daily lint report and maintenance queue, and appends to _log.md; it never changes raw/, KB articles, context packs, or ordinary workspaces.