| name | correct |
| description | Fix misclassified captures. Records corrections to track classification accuracy and improve the system over time. Triggers on: '/correct', 'that should be a task', 'wrong classification', 'reclassify this', 'move this to ideas'. |
Correct Skill
When the owner says a classification was wrong, fix it AND record the error so the system learns. A correction that isn't logged teaches nothing.
Pipeline
Identify item → Confirm correction → Move vault file → (Update mirror, if enabled) → Log to accuracy log → Compile
Step 1: Identify the Item
The owner may reference the item by:
- Title: "the streaming-API thing should be research, not an idea"
- Recency: "that last capture was wrong"
- Path: "vault/ideas/YYYY-MM-DD-some-note.md"
Find the exact file — check vault/index.md and recent captures first (today, then yesterday).
Step 2: Confirm the Correction
Correction:
"Streaming API comparison"
Current: idea (confidence: 0.82) → vault/ideas/
Proposed: research → vault/research/
Confirm?
Step 3: Move the Vault File
- Read the current file
- Update frontmatter:
type: → new type
corrected_from: → old type
corrected_at: → current ISO timestamp
- Move the file to the new type's folder (a move, not a copy)
- Update the
vault/index.md row (section and path)
- Fix any
[[wikilinks]] that pointed at the old location, if your wiki tool resolves by path
Step 4: Update Mirror (OPTIONAL)
Only if integrations.cloud_mirror is enabled in the CLAUDE.md config block and the file has a mirror_page_id: update the mirrored page's type/classification, moving it between databases if needed. Otherwise skip.
Step 5: Update the Accuracy Log
Append to vault/notes/accuracy-log.md (create it with a table header on first use):
| YYYY-MM-DD | idea → research | 0.82 | Streaming API comparison | chat |
Columns: date, original → corrected type, original confidence, item title, capture source.
Step 6: Log + Confirm
Append to vault/log.md:
## [YYYY-MM-DD HH:MM] correct | "Streaming API comparison" idea → research
Then confirm briefly:
Corrected: "Streaming API comparison"
idea → research | moved to vault/research/ | accuracy log updated
Batch Corrections
Multiple corrections at once ("those last 3 captures should all be tasks") → handle in a loop, one accuracy-log row each, summary table at the end.
What Corrections Tell Us
Each correction is a signal. After recording, briefly note a pattern ONLY if there are 3+ corrections of the same kind:
- "This is the 3rd idea → task correction on terse inputs — I'll weigh action verbs more heavily."
Recurring patterns (3+) should be surfaced during /reflect's weekly accuracy review, and the fix (an extra classification rule or example) proposed to the owner before being added to the capture skill's type rules.