Wiki Curate

Wiki Curate — Read-only Audit

Reads the vault wiki and produces a triage report. Never writes. Detects four classes of issue: overlap clusters (fragmentation), tag gaps (untagged concepts + orphan/near-duplicate tags), write-only concepts (zero backrefs from the corpus), and anatomy/lifecycle issues (drafts awaiting curation + curated pages that regressed the concept anatomy). Companion to wiki-grow (which adds new concepts) and wiki-merge (which collapses overlap clusters).

When to Use

• After a wiki-grow run, to check what the new stubs introduced.
• Before deciding which concepts to refine or merge.
• Monthly, to keep the wiki's signal-to-noise ratio high.

When NOT to Use

• For project knowledge folders — those are audited via knowledge-lint.
• For atlas topics — that's atlas-audit.
• For mechanical edits — this skill never writes. Use wiki-merge to collapse overlap clusters and the planned tag-inference pass in wiki-grow to backfill missing tags.

Architecture

scripts/wiki-curate-scan.py
  ├─ reads:  ~/vault/concepts/*.md
  │           ~/vault/atlas/<theme>/*.md         (for backref counts)
  │           ~/vault/books/<slug>/*.md         (for backref counts)
  │           <research-root>/<theme>/<proj>/CLAUDE.md   (for backref counts)
  └─ writes: /tmp/wiki-curate-report.md
             /tmp/wiki-curate.json

The skill is invoked by running the Python orchestrator. No Claude dispatch — the audit is pure deterministic analysis.

What it checks

A. Overlap clusters (the headline feature)

Concept slugs are tokenised (kebab parts, stopwords removed, trailing s stripped for crude singularisation). Any token shared by ≥2 concept slugs forms a cluster.

Within each cluster, the canonical winner is suggested by:

1. Human-curated (auto_generated: false) beats auto-generated.
2. Longer body beats shorter.
3. Alphabetically earliest slug breaks ties.

Pair-level title-Jaccard ≥ 0.40 is flagged as 🔥 high confidence (distinct from "shared keyword" — the slugs are substantively similar).

Example output from a recent run:

### `goodhart` (size 3)
- `goodhart-resistance`  — auto, 3533 chars
- `goodhart-taxonomy`    — auto, 3902 chars
- `goodharts-law`        — curated, 3111 chars  ← canonical (suggested)

Suggested action: wiki-merge goodharts-law goodhart-taxonomy goodhart-resistance

B. Tag coverage

• Untagged concepts — frontmatter has no tags: field. Filters on /concepts?tag=X won't surface them.
• Orphan tags — used by exactly one concept. Either expand or retire.
• Near-duplicate tags — pairs differing only by case / hyphen / trailing s (e.g. ml vs machine-learning, model vs models).

C. Backref health

For each concept, count:

• Incoming wikilinks — [[slug]] references from atlas topic bodies, book chapters, and project CLAUDE.md files.
• Corpus mentions — case-insensitive whole-word slug occurrences in the same corpus.

Concepts with both counts zero are flagged as write-only — they don't surface anywhere. Two remedies:

1. Insert [[slug]] references in the atlas topics that should reference the concept (mechanism A from this session — body wikilinks auto-surface as chips on the topic page).
2. Add the slug to ~/vault/concepts/.denylist and delete the file — accept the concept isn't carrying weight.

D. Anatomy & lifecycle

The status: frontmatter field is the lifecycle axis: draft (no guarantees, may be project-voiced) → curated (conforms to the concept anatomy). Status is read from frontmatter with the same fallback as atlas-workspace (auto_generated: true → draft, else curated) so pages predating the field classify correctly. The anatomy is defined in packages/atlas-vault/schema.md → "Concept File Schema":

• ## In one line — required definition block (machine-anchored).
• ## In my portfolio — the boundary below which project-voiced prose ("the project", "this paper", "we propose", …) is allowed.

Two lists are reported:

• Drafts awaiting curation — every status: draft concept, with a ✓/— table showing which anatomy pieces already exist (In one line, In my portfolio) and how many project-voice phrases sit above the boundary. Higher counts mean more restructuring to do.
• Curated pages violating the anatomy — pages marked curated that regressed: missing the In one line definition, or project-voice prose above the In my portfolio boundary (including a portfolio section under a non-canonical heading). These are the priority fix — a curated page is supposed to conform.

The curation action itself is a Claude-driven restructuring edit (this skill is read-only); see the triage workflow below.

Output

/tmp/wiki-curate-report.md with three sections (A/B/C) and a suggested triage order. Stdout from the script:

Concepts: 14 (6 curated, 8 draft)
Overlap clusters: 0
Drafts awaiting curation: 8
Curated pages violating anatomy: 1
Untagged: 8/14
Write-only (no backrefs): 0
Report: /tmp/wiki-curate-report.md
JSON:   /tmp/wiki-curate.json

The JSON sidecar feeds the planned wiki-merge skill so it can read the suggested winner/fold-in lists without re-running the detection.

Modes

There's no --autonomous flag here — the skill is already pure-read and idempotent. Future cron integration: chained after the Saturday 06:45 wiki-grow run, the audit can land in log/audits/wiki-curate-YYYY-MM-DD.md for a weekly health snapshot.

Triage workflow

After running:

1. Section D regressions first — any curated page flagged as violating the anatomy is a bug: it claims curated but doesn't conform. Fix in place (add ## In one line, or rename/add the ## In my portfolio section and move project-voiced prose under it).
2. Section A — every overlap cluster fragments knowledge. Pick a canonical winner (often the suggestion) and run wiki-merge <winner> <fold-in-slugs> to collapse. Re-run wiki-curate to confirm the cluster is gone.
3. Section D drafts — curate each draft to the anatomy:
   • Extract a 2-4 sentence project-agnostic definition into ## In one line.
   • Keep the cited, stable knowledge in encyclopedia voice.
   • Move all project-voiced prose under ## In my portfolio.
   • Verify references: keys are canonical Paperpile keys (rekey hand-minted lookalikes).
   • Flip status: draft → status: curated in frontmatter and set last_updated.
4. Section B — untagged concepts won't surface in the tag filter. Until the planned wiki-grow tag-inference pass lands, add tags manually for the auto-promoted stubs you care about.
5. Section C — write-only concepts. Decide per-concept: link from a topic, denylist, or accept.

Cross-References