メニュー
Find near-duplicate notes and merge them safely. Use when the user asks to "dedupe the KB", "find duplicate notes", "merge these notes", or after large research sessions that may have re-covered existing ground.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Re-verify time-sensitive claims in KB notes against live sources and apply visible corrections. Use when the user asks to "refresh the KB", "re-verify notes", "check what's gone stale", or on a schedule. Distinct from /kb-review (structural quality) — this checks whether the facts are still true.
Synthesize a coherent narrative explanation of a topic from multiple KB notes. Use when the user asks to "explain X from the KB", wants a structured overview/briefing woven from existing notes, or asks "what's the full picture on X".
Answer a single question and save the answer as an atomic note in the knowledge base. Use when the user poses a discrete factual/conceptual question they want captured for later, says "look this up and save it", "add this to the KB", or "note down the answer to X".
Agentic research loop that deeply investigates a topic via web search and saves findings as atomic notes
Audit the knowledge base for quality issues (orphans, broken links, missing citations, stale notes, tag drift) and suggest improvements. Use when the user asks to "review the KB", "check note quality", "find orphans/broken links/duplicates", or "what needs cleanup".
Search the knowledge base and synthesize a grounded answer from matching notes. Use when the user asks a question the KB likely covers, says "what do we know about X", "search the KB", "search my notes", or "answer from the KB".
| name | kb-dedupe |
| description | Find near-duplicate notes and merge them safely. Use when the user asks to "dedupe the KB", "find duplicate notes", "merge these notes", or after large research sessions that may have re-covered existing ground. |
| user_invocable | true |
| arguments | Optional: --kb <name>, --threshold X (default from config similarity.dedup_threshold), --limit N, or two slugs to merge directly |
Find likely-duplicate note pairs and resolve them: merge, link, or dismiss. The engine handles mechanical rewiring; you make the editorial calls.
Direct merge: If $ARGUMENTS contains two slugs, skip discovery and go to step 4 with that pair.
Discover — Run python3 .kb/kb-index.py dedupe --json [--kb <name>] [--threshold X] [--limit N]. If empty, report and stop.
Classify each pair — Read BOTH notes fully before judging. High cosine similarity has three distinct causes; only one is a duplicate:
related: already (already_linked: true), lean toward intentional.For true duplicates, pick the survivor — Prefer: the note with more inbound links; better sources (primary > secondary); higher epistemic status; better title/slug. When equal, keep the newer.
Merge — For each confirmed pair:
a. First move any unique content from the loser into the survivor (facts, sources, takeaways it lacks — respect the note-format sections). Update the survivor's updated: date.
b. Run python3 .kb/kb-index.py merge <loser> <survivor> — this backs up automatically (.kb/backup.sh), rewires every [[wikilink]]/related:/depends_on: reference across all KBs, unions sources, and marks the loser deprecated_by the survivor.
c. Use --dry-run first if the pair has many inbound links; sanity-check the rewire list.
d. Delete the loser stub only if the user asked for deletion; default is to keep the deprecated stub.
Verify — python3 .kb/kb-index.py build --incremental, then lint <survivor> for each merge. Confirm graph orphans didn't grow.
Log — Append a ## [YYYY-MM-DD] dedupe | <scope> entry to .kb/log.md: pairs reviewed, merges done (loser → survivor), pairs dismissed and why.
Table of pairs: score | verdict (merged / template siblings / intentional variants) | action taken. Lead with merges. Never merge without having read both notes — similarity scores cannot distinguish duplicates from templates.