Prose

Prose

Overview

Prose written or touched by an AI carries predictable tells: throat-clearing openers, fake binary contrasts, inanimate things doing human verbs, runways of negation before the actual point. This skill names the patterns and gives the fix. It applies to anything a reader reads as a sentence: articles, READMEs, landing and UI copy, emails, docs. Not code comments, not commit messages.

The detailed pattern catalog lives in references/banned-patterns.md. The rules below are the ones to keep in your head while drafting.

1. Core rules

• Name the human. Inanimate things don't act. "The team fixed it that week" beats "the complaint becomes a fix." Decisions don't emerge, cultures don't shift, data doesn't tell you anything: a person decides, changes behavior, reads the data. If no specific person fits, use "you" and put the reader in the seat.
• State it directly. No "not X, but Y" contrasts, no setup-and-reversal. "Speed isn't the problem, accuracy is" becomes "Accuracy is the problem." Drop the negation and assert the point.
• Show the specific thing. No vague declaratives. "The implications are significant" names nothing: name the implication. If a sentence calls something important, deep, or structural without showing it, replace it with the specific thing or cut it.
• Cut the runway. No negative listing before the point ("Not a framework. Not a library. A runtime."). State the destination. The reader doesn't need the buildup.
• One idea per sentence. Find the actor, put them at the front, say what they did. Active voice over passive. Vary sentence length so the rhythm isn't metronomic.

2. Quick checks

Run these before delivering a text:

• Passive voice anywhere? Find the actor, make them the subject.
• Inanimate thing doing a human verb ("the decision emerges")? Name the person.
• Sentence starts with a Wh- word (What/When/Why/How)? Restructure, lead with the subject.
• Any "here's what / this / that" throat-clearing? Cut to the point.
• Any "not X, it's Y" contrast? State Y directly.
• Negative listing before the reveal? State the destination.
• Vague declarative ("the implications are significant")? Name the specific thing.
• Narrator-from-a-distance ("Nobody designed this")? Put the reader in the scene.
• Meta-joiners ("the rest of this essay...")? Delete. Let the text move.
• Business jargon (navigate, deep dive, circle back)? Swap for plain words.
• Dramatic fragment ("That's it. That's the fix.")? Use a complete sentence.
• Three consecutive sentences the same length? Break one.

3. French specifics

The repo owner writes content in French and code in English. For French content:

• Never use em dashes (tirets cadratins). This is a hard project preference for French content: use commas, periods, or parentheses instead. In English, em dashes are fine sparingly.
• Concrete, descriptive titles, not abstract or "creative" ones. "Opacity seule" over "Le crossfade". A title should say what the section is about.
• Don't over-explain after a code snippet or example. If the code or example speaks for itself, a short sentence is enough. No paragraph restating what the reader just saw.

4. Full pattern list

For the full editing pass, with every pattern, a concrete example, and its fix, go to references/banned-patterns.md. It carries the catalog as tables: throat-clearing, emphasis crutches, the business jargon replacement table, meta-commentary, false intimacy, telling-instead-of-showing, vague declaratives, the eleven fake binary contrasts, negative listing, dramatic fragmentation, rhetorical setups, false agency, narrator-from-a-distance, passive voice, and sentence-starter / rhythm fixes. Match a pattern, apply the fix, move on.