| name | deslopify |
| description | Deslopifies AI-sounding prose and code diffs. Use when the user asks to make text sound human, remove ChatGPT-style filler, rewrite without fluff, clean up AI comments, or strip restatement comments from a diff without changing code. |
Deslopify
Strip the patterns that mark text as AI-generated. Preserve all factual content and the author's intended tone โ change the delivery, not the message.
The Slop Patterns
Scan for these before rewriting. More slop patterns justify a stronger edit; mostly-clean text gets a light pass.
Opening filler (cut 80% of the time โ start with the actual point):
- "In today's fast-paced world..."
- "It's important to note that..."
- "As we navigate the complexities of..."
Hollow transitions:
- "Furthermore,", "Moreover,", "Additionally," โ usually just "Also," or delete
- "In conclusion," โ end without announcing you're ending
- "It's worth noting that..." โ just say the thing
Emphasis inflation:
- "Importantly,", "Crucially,", "Significantly," โ earn emphasis with content
- "It goes without saying..." โ then don't say it
AI vocabulary (replace with the plain word):
- "leverage" โ "use"
- "utilize" โ "use"
- "delve into" โ "look at" or "explore"
- "in terms of" โ restructure the sentence
- "moving forward" โ cut
- "at the end of the day" โ cut
Excessive hedging:
- "It seems that", "One might argue", "It could be said" โ commit to the claim or cut it
Nominalization (convert to verb):
- "make a decision" โ "decide"
- "provide assistance" โ "help"
- "conduct an analysis" โ "analyze"
- "have a discussion" โ "discuss"
Bullet lists where prose would be clearer: not every list needs to be a list. Three related sentences often read better as a paragraph.
En-dash used as em-dash: โ should be โ (or just a comma)
What Good Looks Like
- Short sentences land hard. Longer ones build context. Vary them.
- Active voice: "We shipped X" not "X was shipped by the team"
- Specific beats general: "increased conversion by 18%" not "improved results"
- The first sentence is the most important โ if it's filler, cut it
Process
- Read the full text.
- Identify which slop patterns appear; mention only the main classes in the note.
- Rewrite: cut filler, verb-ify nominalizations, activate passive constructions.
- Present the revised version or cleaned diff.
- Add a 2โ3 line note on the key changes made, not a list of every micro-change.
Output Criteria
Prose mode: return the revised text plus a 2โ3 line note naming the main classes of edits. Do not list every micro-change.
Code comment mode: return a cleaned diff or edit only comments, then report what was removed. Be surgical: no code, control-flow, formatting, or behavior changes.
Code Comment Mode
When the input is a diff or the user asks to remove AI comments from code:
What to remove:
- Comments restating what code clearly does (
// increment counter above counter++)
- Overly verbose explanations of obvious code
- Generic placeholders (
// TODO: implement, // Handle error)
- Comments with AI phrasing ("This function...", "The following code...")
- Inline docs inconsistent with the file's existing comment style
What to keep:
- Doc comments for public APIs (JSDoc, docstrings)
- Comments explaining why (business logic, workarounds, edge cases)
- Comments matching the file's existing style and density
- License headers and attribution
Hard constraint: DO NOT modify any code. Only remove comments. No refactoring, no style changes, no removing try/catch or null checks. If you spot code slop, note it in the summary but leave it.
Hard Rules (Prose Mode)
- Never change factual content โ only how it's expressed
- Never add new claims or examples not in the original
- Preserve the author's stance and intended audience
- If the text is mostly fine, say so โ don't invent problems to fix