Knowledge Wiki Skill

knowledge/ is a visible knowledge asset made of explicit in-repo .md files (committed and shared). It is distinct from Claude Code's memory/ — memory is cross-session personal notes, this is competition knowledge that gets committed.

For the design philosophy and page conventions, see knowledge/README.md. This skill mechanizes its operation.

Subcommands (dispatch on the first token of $ARGUMENTS)

blank: status check

1. Read knowledge/INDEX.md (do not read all pages)
2. Summarize page counts per category and the status distribution
3. Propose knowledge gaps: insights present in recent daily reports but not yet promoted; pages stuck at idea/testing

add <topic or insight>: distill and promote

1. Identify the target insight (from $ARGUMENTS, or the latest daily report / SESSION_NOTES)
2. Dedup check: scan INDEX; if a page on the same topic exists, update it instead of creating a new one
3. Decide the category (technique / data / error / decision / add a new one if needed)
4. Create or update a page following the _template.md frontmatter:
   • id matches the filename (kebab-case)
   • always record provenance in source (daily-report date / exp folder)
   • set status appropriately (validated if confirmed, rejected if discarded)
   • fill impact with numbers (CV/LB delta) if available
   • cross-link related pages via links and [[id]] in the body
5. Distill the body: takeaway → evidence → how to apply → caveats. Don't paste raw daily-report logs
6. Append one line to INDEX.md (under the right category, with a hook)

find <query>: retrieve and load

1. Read INDEX.md, pick 2-3 pages relevant to the query from their hooks
2. Load only those pages and summarize
3. Follow related [[links]] or suggest doing so

promote: harvest from the daily report

1. Read the latest daily_reports/*.md
2. Extract insights that have become "reusable conclusions" (exclude one-off work logs)
3. Cross-check against existing pages; list the un-promoted ones as page proposals
4. With user confirmation, run the add equivalent

consolidate: cleanup (roughly weekly)

1. Bulk-load all page frontmatter (load bodies only as needed)
2. Detect and propose:
   • duplicates: multiple pages on the same topic → merge
   • contradictions: conflicting conclusions → confirm which is newer/correct
   • staleness: pages whose premise changed → mark stale
   • orphans: pages no one links to → add links
3. Regenerate INDEX.md (update page count and last-consolidated date)

Principles

• Daily report first, knowledge second. Don't write straight to knowledge skipping flow (provenance is lost)
• Stay atomic. One topic per page. Split when it grows
• Keep rejected insights (rejected). "Didn't work" is valuable knowledge too
• Keep INDEX current. Every new page gets one line in INDEX
• When recalling, go INDEX → only the relevant page. Don't scan every page