SKILL.md
Copy Editing
Editing is enhancement, not rewriting. Run focused passes, one dimension at a time, because a single unfocused review misses what a targeted sweep catches. Preserve the author's voice and core message.
Before You Edit
Check for a product marketing context file (.agents/product-marketing.md or .claude/product-marketing.md). If it exists, read it and use its brand voice and customer language. Then confirm: what is this copy's goal (awareness, conversion, retention), what action should readers take, and is this a polish job or a content refresh.
The Seven Sweeps
Run them in order. After each sweep, re-check that earlier sweeps still hold, edits in one dimension can break another.
| # | Sweep | The question | Fail signs |
|---|---|---|---|
| 1 | Clarity | Can a stranger understand it in one read? | Jargon, ambiguous pronouns, buried points |
| 2 | Voice | Does it sound like one person throughout? | Casual open, corporate middle, random tech-speak |
| 3 | So What | Does every claim answer "why should I care"? | Features with no "which means..." bridge |
| 4 | Prove It | Is every claim backed? | "Industry-leading", "trusted by thousands", no names or numbers |
| 5 | Specificity | Is it concrete enough to believe? | "Save time" instead of "save 4 hours a week" |
| 6 | Emotion | Does the reader feel the pain or the win? | Flat, informational, nothing at stake |
| 7 | Zero Risk | Is every barrier to action removed? | Objections unaddressed, no guarantee, vague next step |
Per sweep: mark problems first, then propose specific edits, then verify intent survived. Do not fix mid-read.
For the quick-pass word lists (cut words, weak-to-strong swaps), the expert panel scoring method for high-stakes copy, and the content refresh checklist: read references/checks.md when you finish the sweeps or when the user asks for a fast review instead of the full seven.
Working Collaboratively
- Run a sweep, present findings with reasons.
- Propose specific edits, never just "this is weak".
- Let the author decide. It is their copy.
- Re-verify earlier sweeps after each round.
- Stop when a full pass finds nothing new.
Verification
Do a final cold read of the edited copy top to bottom. Expect: no sentence you stumble on, every claim carries proof or a specific number, exactly one clear next step for the reader. If any sentence stumbles, it failed Clarity, run sweep 1 on that section again before delivering.
Good vs Bad
Judgment call: a claim with no proof.
Bad edit: "Customers love us" becomes "Customers really love us". Intensifier added, nothing proven.
Good edit: "Customers love us" becomes "Rated 4.8 from 312 reviews, and 40% of new customers come from referrals." Claim now carries its own evidence.
Footguns
- Rewriting instead of editing. If your edit changes the argument, you switched skills. Fix: flag it and offer to run copywriting instead, or constrain edits to the existing message.
- Fixing everything in one pass. You will catch clarity issues and miss proof gaps. Fix: one dimension per pass, no exceptions.
- Later sweeps breaking earlier ones. Adding emotion in sweep 6 often adds vagueness. Fix: after every sweep, spot-check the previous ones.
- Polishing copy that should be retired. Outdated stats and dead examples need a refresh decision, not line edits. Fix: use the refresh vs rewrite matrix in references/checks.md.
Completion Checklist
- [ ] Context file checked, goal and audience confirmed
- [ ] All seven sweeps run in order, or quick-pass explicitly agreed with the user
- [ ] Every proposed edit has a stated reason
- [ ] Core message and author voice preserved
- [ ] Final cold read passed
Any box unchecked: not done. Fix or say so.