The posts on this site start as agent drafts that I edit. The drafts come with the awkward phrasing AI is known for. At first I fixed each case by hand, and the next post carried the same disease. I was fixing sentences, not habits.
So I changed the setup. I no longer fix awkward sentences. I only flag them. The agent takes the flag, names the pattern, and stores it as a rule in a skill file. From the next post on, the rule applies at draft time. One flag becomes a permanent rule instead of a one-time edit.
A single skill file is enough to start
A Claude Code skill is one markdown file. Put it at .claude/skills/prose-polish/SKILL.md and the agent loads it when polishing a draft, following its procedure and rules.
Ours has four parts: the procedure (read the whole draft, fix rule violations, check rhythm by reading aloud, run a community skill as a second pass, report changes), a list of Korean tells, a list of English tells, and the voice to protect. That last part matters. A corrector that flattens your voice is worse than none, so the file also says what to keep: short punch sentences, honest notes about limitations.
