Notion cleanup

/notion-cleanup

Read-only audit + opt-in batch fixes. Default behaviour is reporting; the user must approve each fix category before any write happens.

Inputs

• Database URL or ID (required).
• Stale threshold (optional, default 90 days). Rows with last_edited_time older than this are flagged as stale.
• Required-field list (optional). User can specify "rows must have Cost and Status filled". If omitted, infer from typical patterns (Status filled, Title non-empty, primary numeric field non-empty).

Steps

1. Read. notion-fetch the DB schema + use the row-enumeration workaround (notion-search with single-letter queries "a" "e" "o" "s", union results, fetch each by ID for full properties). Cap at 500 rows for performance; for larger DBs, scope to a filter first.
2. Run checks. For each row:
   • Stale? last_edited_time < (now - threshold).
   • Required blanks? Any required property is empty.
   • Mistyped values? A text property contains values that look like they should be a number / date / URL (heuristic: 95%+ of values match a single non-text pattern).
   • Duplicate titles? Two or more rows share the same title (case-insensitive).
   • Status mismatch? A row marked "Active" but no edit in 90+ days.
   • Orphan select options? Options defined in schema but no row uses them.
   • Workstream mismatches? Multi-select rows where one tag contradicts another (e.g. "Cancelled" status + "Active campaign" workstream).
3. Triage report. Group findings by category, with row counts + sample row URLs (first 3 per category). Sort categories by impact (mistyped column > required blanks > duplicates > stale > orphan options).
4. Confirm gate per category. For each category, ask: "Fix these N rows? (y / skip / show all)". Do not bundle approvals.
5. Apply fixes per approval:
   • Stale "Active" rows → propose flipping to "Paused" or "Cancelled". Show each, never blanket-update.
   • Mistyped column → propose ALTER COLUMN ... SET <type> (with the conversion preview from /notion-tracker rules).
   • Duplicates → propose merging or marking the older row "Archived". Never delete.
   • Required blanks → list them; never auto-fill values, just surface for human entry.
   • Orphan options → propose ALTER COLUMN to drop them; warn this is one-way.
6. Confirm. Print summary of fixes applied + categories deferred.

Decision rules

• Read-only by default. This skill's value is in the report; mutations are opt-in per category.
• Never blanket-update. Even if user says "fix everything", iterate per category, show the proposed change set, get explicit y on each.
• Never delete rows. Worst-case mutation is changing Status to "Archived". Deletion is a manual user action.
• Type conversion is irreversible if values get lost. Always preview the lossy values before the ALTER.

Output

• Triage report in conversation.
• Per-category fix summary after each approval.
• Optional: append findings to Admin/<tracker-name>-cleanup-log.md in the vault so next audit can compare.

Anti-patterns

• Don't audit then immediately fix without showing the report. The report IS the deliverable for most runs.
• Don't propose "AI-fixing" blank required fields by inference. That's fabrication. List them and stop.
• Don't audit cross-database relations (rollups, relations) in v1 — too brittle, scope expansion.

Hard-won gotchas

• Stale threshold should match the DB's natural cadence. Marketing campaign DBs may be stale at 7 days; OPEX DBs at 90+ days. Default 90 but state it loudly and ask if it should differ.
• last_edited_time reflects ANY edit including bulk schema changes. If user just renamed a property, all rows look "fresh" — caveat the staleness check.