1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
---
name: editing-prose
description: Use when writing or editing prose longer than a paragraph to catch AI writing tropes — patterns that flag text as LLM-generated rather than human-written.
---

# Editing Prose

Catches AI writing tropes that other writing skills miss. Sources:
[tropes.fyi](https://tropes.fyi) and
[Modeltell](https://github.com/thirdshiftlab/modeltell).

Complements `technical-writing` (structure, terminology, sentence-level
clarity) and `diataxis` (document classification). Those skills cover good writing in
general; this one covers the specific patterns LLMs reach for that flag
text as AI-generated.

## When not to use

- Code comments (different conventions; brevity matters more)
- Commit messages (use `describing-changes`)
- Short replies, single-sentence answers, terse status updates
- Code itself

## Workflow

1. Read [`ai-tropes.md`](ai-tropes.md) for the full catalog, with
   examples for each pattern. Each tropes.fyi entry is tagged with a
   category (word choice, sentence structure, paragraph structure,
   tone, formatting, or composition) and a trend (`new`, `rising`,
   `consistent`, or `fading`).
2. Scan the prose for each trope category, looking for repetition or
   stacking — one instance of a pattern is usually fine; multiple is
   the tell.
3. Edit in place. Replace each flagged pattern with plain, varied
   prose.
4. Re-read the edited version end-to-end. AI tropes often hide in
   transitions and conclusions; a second pass catches what the
   category-by-category scan missed.

## The core principle

Don't mechanically eliminate every em dash or every "is" → "serves
as" — that produces stilted, evasive writing in the other direction.
Look for *concentration* and *stacking*. A single em dash in a
1000-word piece is invisible; ten is the tell.

## High-signal tropes (check these first)

These are the strongest tells. If you see them, edit them:

- Negative parallelism: "It's not X — it's Y" / "not because X, but
  because Y"
- "Not X. Not Y. Just Z." countdown
- "The X? A Y." self-posed rhetorical questions
- Em-dash density (more than ~3 per 1000 words is suspicious)
- Bold-first bullets (every list item starts with `**Word**:`)
- "Delve", "tapestry", "landscape", "leverage" as a verb
- "Serves as"/"stands as"/"represents" instead of "is"
- Tricolon stacking (rule of three repeated three times in a row)
- "Here's the kicker"/"Here's the thing" false suspense
- Unicode arrows (`→`) and smart quotes in plain text contexts
- "Whether you're X or Y" pseudo-inclusive openers
- "In today's [adjective] landscape/world/era" openers