name: humanize-prose
description: Strip AI writing tells from prose you draft or clean: PR descriptions, Slack messages, docs, emails, commit messages. Use whenever the user asks to "humanize" or "deslop" text, or runs /humanize-prose. Also applies implicitly: any prose you draft for the user must already satisfy these rules.
Strip AI writing tells out of prose, whether cleaning existing text or generating it. Edit in place, change as little as possible, preserve meaning. When a change would alter meaning, leave it and flag it.
Cut, across the whole text, not just the first instance of each:
- Throat-clearing openers: "In this PR, I've...", "This change introduces...", "I wanted to...".
- Hedging filler: "it's worth noting", "it's important to", "generally speaking", "as you may know".
- Marketing adjectives and trendy intensifiers: robust, seamless, comprehensive, powerful, leverage, streamline, delve, load-bearing.
- Rule-of-three padding and restating the obvious.
- Connective tissue that links sentences without adding content: "That said", "With that in mind", "To that end", "As such", "Moreover", "Furthermore", "Additionally", "It follows that", "What this means is...". Drop the phrase or replace it with a full stop.
- Duplicated information: a point already made earlier in different words, a closing line that repeats the opening, a sentence that restates its own heading. Keep it once.
- Contradictory information: two statements in the same text that conflict (a number, a claim, or a recommendation stated one way then another). Keep the version that is correct and flag the contradiction; do not silently pick one if you cannot tell which is right.
- Bullet bloat where two sentences of prose are tighter; bold-everything formatting.
- Summary-of-the-summary closers: "In conclusion...", "Overall, this...".
- Em-dashes and double hyphens (use a comma, colon, or full stop instead).
- False agency: passive constructions or abstract/inanimate subjects performing human actions ("the complaint becomes a fix", "mistakes were made", "the decision emerges"). Name the actor.
- Binary-contrast crutch: manufactured "not X, it's Y" framing standing in for a direct claim.
- Vague declaratives and hollow absolutes: unsupported generalizations ("the implications are significant") and lazy extremes ("always", "never", "every") doing the work a specific claim should.
Keep it concrete: what changed, why, what to check. Match the channel (a PR body is not a Slack one-liner).
slop: In this PR, I've made some changes to refactor the billing service. It's
worth noting this is a fairly comprehensive update that should make the
code more robust and maintainable.
clean: Refactors the billing service: extracts charge calculation into
ChargeCalculator and removes the duplicated rounding logic.
Commit messages still follow Conventional Commits 1.0.
Output
Make edits silently, without narrating each change. End with a 1-3 sentence summary of what categories you cut and anything you left in place and flagged (a contradiction you could not resolve). Nothing else.
