scribe

Scribe

A Claude Code plugin for knowledge capture. Turns unstructured input — spoken or typed — along with existing documents and scattered repo files into structured, indexed, cross-referenced notes.

Skills

Skill	Command	What it does
Dictation	/scribe:dictation	Organize raw ideas (spoken or typed) into structured, indexed notes
Ingest	/scribe:ingest [path]	Bulk import .md, .docx, .txt, .pdf files into structured notes
Discover	/scribe:discover	Scan a repo for documents, present a summary, hand off to ingest

Install

# Option 1: Load for a single session
claude --plugin-dir /path/to/scribe

# Option 2: Add as a local marketplace, then install
/plugin marketplace add /path/to/scribe
/plugin install scribe@scribe

How it works

All notes go into claude-notes/ in your project directory to avoid colliding with existing files. Each note follows a consistent schema:

claude-notes/
|- INDEX.md          # Searchable table of all notes
|- topic-one.md      # Structured note with Context, Idea, Impact, Related, Next Steps
|- topic-two.md
\- ...

Notes are additive — new content gets timestamped and appended, never overwrites. Contradictions are preserved with timestamps so the thinking evolution is visible.

Merge heuristic

When new content arrives, scribe checks the existing index and decides:

• Merge when content extends, refines, or adds a dimension to an existing note
• Create new when content is conceptually independent, even if it shares tags

User profile

Scribe learns your communication patterns over time. The runtime profile lives at ~/.scribe/user-profile.md (private, never committed to project repos). The references/user-profile.md files in this repo are public examples showing the profile format.

Session tracking

Scribe tracks sessions to enable evidence-based profile learning. Each skill invocation writes a session record to ~/.scribe/sessions/ (one file per session, write-once, never modified). Session files use the naming format {YYYYMMDDTHHMMSS}-{skill}-{5-random}.md.

Evidence counting works by grepping across session files for canonical pattern slugs. A candidate pattern is promoted to the main profile sections after appearing in 2+ sessions (or immediately if the user states it explicitly). A reconciliation pass at session start catches any promotions lost to concurrent writes or interrupted sessions.

See skills/shared/references/session-tracking.md for the full protocol.

Note schema

Every note uses this structure:

# Topic Title

**Created:** YYYY-MM-DD
**Updated:** YYYY-MM-DD
**Tags:** #tag1 #tag2
**Source:** dictation | filename.ext | discovered

## Context
## Idea
## Impact
## Related
## Next Steps
## Source Log

See skills/dictation/references/note-schema.md for the full specification.

License

MIT