| name | clean-correction |
| description | When overwriting existing content, write only the intended replacement — never leave traces of the correction process. |
| argument-hint | Target artifact and intended replacement |
Clean Correction
When the operator redirects or corrects an earlier draft, the durable artifact must contain only the final intended content — not a record of how we got there.
The Rule
Write only what the artifact should say, not what it used to say or how it was changed.
- Do not include
A (not B), A, not B, A instead of B, A as corrected, or similar traces.
- Do not paraphrase the operator's correction back into the artifact.
- Do not leave stubs of the previous version behind.
Procedure
- Read the target before writing. Know what is being replaced.
- Write the replacement directly. No commentary about the edit.
- Verify the result contains only the intended content.
When To Pause
Pause and ask for clarification instead of guessing when:
- the target boundary is unclear (which part of the file is being replaced?)
- the replacement content is still implicit or underspecified
- the change is large enough that a quick summary of what is going away would help
Edge Cases
- Fuzzy corrections ("make this more casual", "tighten this up"): restate the concrete replacement before writing. Do not just edit and hope.
- Full rewrites: write the new version. Do not preserve fragments of the old one to show the transition.
- Deletions: delete cleanly. Do not replace deleted content with
// removed or [empty].
- Fast-path requests ("just do it"): write cleanly anyway. The discipline is in the artifact, not the conversation.
What This Skill Does Not Do
- It does not classify edits as additive or destructive.
- It does not require an explicit confirmation prompt before writing.
- It does not generate diffs for the operator to review.
The point is clean artifacts, not procedural ceremony.
