| name | reindex-memory |
| description | Rebuilds INDEX.md files across craft/ and experience/. Run after note-taking, pattern-promotion, character-diff application, or by the session-end hook for safety. |
reindex-memory
Walks the markdown library and rebuilds index files. Light-touch, fast, deterministic. No vector indexing — just markdown link tables.
When to invoke
- Triggered by
take-craft-note skill after each note write
- Triggered by
promote-pattern skill after each promotion run
- Triggered by
apply-approved-diff skill after each character change
- Auto-fired by the
session-end-reflect.sh hook at session end (safety reindex)
- Manually invokable for debugging
What it builds
craft/INDEX.md
Top-level navigation across the library plus a brief recent-activity summary.
craft/notes/INDEX.md
Table of all notes with date, slug, tags, status. Promoted notes get a promoted-to: column linking to the pattern.
craft/patterns/INDEX.md
Table of all patterns with slug, promoted date, confidence, status, last-reaffirmed timestamp.
craft/canon/INDEX.md
Table of all canon entries with slug and one-line description from each file's first paragraph.
craft/exemplars/INDEX.md
Table with slug, persona-targeted, rhetorical-move, source. Used by drafter to query exemplars by persona+move.
craft/anti-canon/INDEX.md
Table with slug, primary-failure, persona-targeted. Used by witness for anti-canon lookups.
experience/prospects/INDEX.md
Table with lead-id, persona, industry, campaign, register-used, witness-verdict, outcome (when known), date.
experience/campaigns/INDEX.md
Table with campaign-id, date, sender, service-line, total leads, drafted count, skipped count, flagged count.
Workflow
1. Walk craft/
For each subdirectory, read frontmatter from each markdown file. Build the appropriate index table from the frontmatter fields.
2. Walk experience/
For each prospect folder and campaign folder, read the relevant frontmatter (from final.md, output.json, witness-feedback.md, brief.md).
3. Write index files
Each INDEX.md is replaced wholesale (not appended). The schema for each index is documented in the relevant directory's README.
4. Update craft/INDEX.md last-reindex timestamp
The top-level INDEX file's Recent activity section gets updated with current counts and the reindex timestamp.
5. Surface
Report to the user (or the calling skill):
- How many files were indexed across each subdirectory
- Any anomalies detected (files missing required frontmatter, broken links, orphaned references)
Anomaly handling
If a file lacks required frontmatter, log it but do not fail the reindex. Continue and report the anomaly so the human can repair it. The reindex itself should be robust — a broken file in one corner of the library should not prevent the rest of the library from being indexed.
What this skill does not do
- It does not edit content files. Only index files are written.
- It does not delete files. If a referenced file no longer exists, the index entry is omitted, but the original file is left alone.
- It does not perform any semantic operation (no clustering, no summarization, no inference). Indexing is purely structural.
Performance
The library is small (hundreds of files at most in Phase 1). The reindex should complete in under 5 seconds. If it is slower, investigate — the operation is fundamentally fast.
Why no vector indexing
Phase 1 deliberately avoids embeddings. Markdown indexes plus Grep/Glob over the file system are faster, more interpretable, and more debuggable than vector search at this scale. If retrieval becomes the bottleneck (it will not in Phase 1 with hundreds of files), Phase 2 can revisit.
