scribe-kb

name

scribe-kb

description

Read, write, and search a scribe-managed knowledge base (markdown vault with frontmatter conventions, wikilinks, and qmd hybrid search). Use when the user's project has a scribe.yaml (KB root) or a .claude/<kb_name>/ drop-file directory (consumer side), or when they mention scribe, scriptorium, qmd, drop files, or "my KB". Covers frontmatter schema, wikilink syntax, drop-file pattern, search via qmd, and directory taxonomy.

scribe knowledge base — agent skill

A scribe-managed KB is a local-first markdown vault. Articles live in topic directories under a KB root, every article carries YAML frontmatter, links between articles are [[Wikilinks]], and search runs through qmd (BM25 + vector hybrid). A daemon scribe runs on cron to extract from project repos, mine Claude Code sessions, capture iMessage URLs, absorb queued URLs, lint frontmatter, and reindex qmd.

This skill teaches a Claude Code / Codex / OpenCode session how to interact with such a KB without breaking the conventions the daemon expects.

When to use this skill

Trigger on any of:

• You're working in the KB root — there's a scribe.yaml in the project root, or directories named wiki/, decisions/, solutions/, patterns/, research/, ideas/, tools/, people/, projects/, raw/articles/.
• You're working in a non-KB project that drops files into the KB — there's a .claude/<some_kb_name>/ directory with markdown drop files, or the user's CLAUDE.md mentions a personal KB and a drop-file convention.
• The user asks about: their KB, scriptorium, qmd, drop files, "what do I know about X", "have I done this before", "save this for later", or any of the article types above.

Operations cheat sheet

What	Command (qmd / shell)
Find articles by topic	qmd query "<natural-language question>" (or use the qmd MCP tool when present)
Exact-keyword search	qmd search "<keywords>"
Read a specific article	Read tool with the absolute path
Read a section	scribe sections get <article> <id> (Phase 5A)
Write a new wiki article	Edit/Write tool — see references/FRONTMATTER.md
File a drop from another project	scribe drop --title ... --type ... --domain ... — see references/DROP_FILES.md
Validate frontmatter before commit	scribe validate <file> or scribe lint --changed
List sections in an article	scribe sections list <article>

Workflow: file a drop file from another project

When you produce reusable knowledge in a non-KB project that should travel to the user's KB, write a drop file:

1. Pick a path: .claude/<kb_name>/YYYY-MM-DD-{slug}.md in the current project root. <kb_name> is the value the user has used in their CLAUDE.md (often scriptorium).
2. Run scribe drop with the required flags (--title, --type, --domain, --body) — or hand-write the frontmatter per references/DROP_FILES.md (scriptorium: true, action, title, type, domain, tags) if scribe isn't installed.
3. Write the body — what's reusable, what would teach a future session this insight in 30 seconds.
4. Tell the user what you filed and why. Don't fabricate drop files for trivial facts.

scribe sync running on the user's cron will absorb the drop file automatically — you don't need to do anything else.

Workflow: write a new wiki article

When the user asks you to write inside the KB itself (not a drop file):

1. Pick the directory by article type (see references/STRUCTURE.md). A reusable solution → solutions/. A long-lived decision → decisions/. A research deep-dive → research/. A pattern that applies across projects → patterns/.
2. Write valid frontmatter — required keys: title, type, created, updated, domain, confidence, tags, related, sources. See references/FRONTMATTER.md.
3. Cross-link with [[Article Title]] to existing articles. Search first with qmd; don't invent neighbors.
4. Commit with a clear message. The user's pre-commit hook runs scribe lint.

Workflow: search before recommending

Before recommending a tool, library, decision, or pattern, search the KB:

qmd query "<the thing> verdict"
qmd query "<the thing> alternatives"
qmd query "<the problem> decision reasoning"

Past tool evaluations live in tools/ with verdict: use | evaluate | skip. Past decisions live in decisions/ with full context. Don't suggest something already rejected.

What NOT to do

• Don't write directly to the KB from a non-KB project. Use a drop file. The user's daemon owns absorbs.
• Don't invent frontmatter fields outside the closed set. Lint will reject typos.
• Don't bypass scribe lint when the user asks you to commit. The pre-commit hook is the canonical validator.
• Don't cd into the KB root before searching with qmd — qmd uses absolute paths and works from any directory. CD-ing breaks the user's flow.
• Don't write rolling-file contents into article shapes. learnings.md, decisions-log.md are append-only logs; ordinary articles are paragraph-shaped.

References

• FRONTMATTER.md — required and optional frontmatter keys per article type
• WIKILINKS.md — link syntax (basic, aliased, section, block-anchor)
• STRUCTURE.md — directory taxonomy and what belongs where
• DROP_FILES.md — the cross-project drop-file pattern in detail
• QUERY.md — qmd query patterns: lex, vec, hyde, when to use which
• COMPAT.md — Obsidian / Logseq vault compatibility notes