AI Tells — Surface Scan
A near-mechanical pass that strips the obvious AI fingerprints from prose. Runs before any prose artifact is presented as final.
When This Runs
Auto-engages whenever the task produces prose-as-artifact:
- READMEs, CONTRIBUTING, CHANGELOG
- Technical documentation under
docs/
- PR descriptions, PR comments, commit message bodies
- Tutorials, guides, blog posts, release notes
- Any markdown intended for human readers
Skip for: code comments (different rules), inline chat replies, structured data (YAML/JSON), or anything pure-config.
How To Use
After drafting prose, scan it once against each list below. Every match is a rewrite, not a debate. The lists are deliberately short and high-leverage — if a tell isn't here, it belongs in ai-tells-review.
Banned Punctuation
- Em dashes connecting clauses. Replace with period, comma, colon, or parentheses depending on the relationship. The em dash is the single most diagnostic AI tell in prose.
- Smart quotes when the source is plain ASCII. Mixing curly and straight quotes in a doc is a copy-paste fingerprint.
- Ellipses for dramatic pause in technical writing (
it's...complicated). Reads as performance.
Banned Sentence Openers
- "Importantly," "Notably," "Crucially," "Interestingly"
- "Furthermore," "Moreover," "Additionally"
- "It's worth noting that…" / "It bears mentioning…"
- "In essence," "Ultimately," "At its core"
- "In conclusion," "To summarize," "In summary"
- "Great question!" or any sycophantic acknowledgement
If the sentence after the opener stands on its own, delete the opener. It almost always does.
Banned Vocabulary
These words appear far more often in machine-generated prose than human prose. Each has a fix-direction.
| Word / Phrase |
Why it's a tell |
Fix direction |
delve |
Diagnostic on its own |
use look at, examine, or just delete |
leverage (verb) |
Corporate-AI fingerprint |
use use |
robust, comprehensive, seamless, streamlined |
Empty positive adjectives |
name the specific property |
cutting-edge, innovative, state-of-the-art |
Marketing puff |
describe what it actually does |
elegantly, seamlessly, gracefully |
Empty positive adverbs |
delete the adverb |
modulo (in non-math prose) |
Claude tic |
use except for or aside from |
stakeholders |
Generic abstraction |
name the people or roles |
passionate about, excited to, thrilled to |
Performative enthusiasm |
state what was done |
it's worth noting |
Throat-clearing |
delete; state the thing |
at the end of the day |
Verbal filler |
delete |
move the needle, circle back, align on |
Corporate cliche |
name the actual action |
Banned Structural Reflexes (Surface Layer)
These are mechanical enough to catch on a scan; deeper rhythm/structure tells live in ai-tells-review.
- Tricolon reflex. Three-item lists by default ("fast, reliable, and scalable"). If two items would do, use two. If five, use five.
- "This is not X — it's Y" contrastive bold. The mirrored "Why not pure JSON: / Why not pure scriptblocks:" pattern. Delete the contrast or merge into a single sentence.
- Decision-stamp formality.
**Decided 2026-05-18:** style headers in personal notes. Real humans rarely formalize their own decisions like a court ruling.
- Self-narration. Sentences that comment on what the doc is doing: "That's the dashboard's reason for existing." / "(this is where we start)". Delete and let the doc do the thing instead of describing itself.
- Compliment sandwich. Soften → criticize → soften. State the criticism directly.
- Three consecutive sentences saying the same thing in different words. Delete two of them.
Banned Visual Reflexes
- ASCII art boxes for layered diagrams. When the diagram is more than two boxes, use mermaid or just describe the layers in prose.
- Italicized rhetorical emphasis. missing, separately, on this dashboard. Italics are for terms-of-art and titles, not tone. One per long doc, max.
- Bold for emphasis on every paragraph. Bold should mark scannable structure (headers, key terms), not feeling.
Coined-Phrase Tic Check
Before submitting, scan for any minted phrase repeated 3+ times across the doc. Examples seen in real artifacts: "proves the framework," "earns its place," "5-second glance." Humans usually rephrase the second or third use; models don't. Rephrase all but the first occurrence.
What This Skill Does Not Do
- Catch rhythm patterns (uniform sentence length, paragraph-level parallelism, gerund cascade)
- Catch hedging cadence and manufactured both-sidesing
- Catch aphoristic closers and summary-of-summary endings
- Replace voice skills like
nick-voice — this subtracts machine fingerprints, it doesn't add personality
For all of the above, use ai-tells-review.
Output Shape
The scan is invisible by default — it runs before the prose is presented and the prose comes out clean. If a tell can't be cleanly fixed (e.g. the contrastive-bold pattern is load-bearing for the argument), flag it inline:
NOTE: kept "X — Y" construction at line 42; the contrast carries the argument and rewriting flattens it.
This gives the user one place to check whether to push back on a specific judgment.
1---2name: ai-tells-scan3description: Use when writing tasks produce prose artifacts (READMEs, docs, PR text, tutorials, guides, posts, release notes) and need a fast surface scan for AI tells. Auto-engages on prose writing tasks; checks banned punctuation, stock vocabulary, and mechanical patterns. For the structural judgment pass on rhythm and rhetorical reflexes, use ai-tells-review. Not a replacement for voice skills like nick-voice; this removes machine fingerprints, it does not add a target voice.4---5
6# AI Tells — Surface Scan
7
8A near-mechanical pass that strips the obvious AI fingerprints from prose. Runs before any prose artifact is presented as final.
9
10## When This Runs
11
12Auto-engages whenever the task produces prose-as-artifact:
13
14- READMEs, CONTRIBUTING, CHANGELOG
15- Technical documentation under `docs/`
16- PR descriptions, PR comments, commit message bodies
17- Tutorials, guides, blog posts, release notes
18- Any markdown intended for human readers
19
20Skip for: code comments (different rules), inline chat replies, structured data (YAML/JSON), or anything pure-config.
21
22## How To Use
23
24After drafting prose, scan it once against each list below. Every match is a rewrite, not a debate. The lists are deliberately short and high-leverage — if a tell isn't here, it belongs in `ai-tells-review`.
25
26---
27
28## Banned Punctuation
29
30- **Em dashes connecting clauses.** Replace with period, comma, colon, or parentheses depending on the relationship. The em dash is the single most diagnostic AI tell in prose.
31- **Smart quotes when the source is plain ASCII.** Mixing curly and straight quotes in a doc is a copy-paste fingerprint.
32- **Ellipses for dramatic pause** in technical writing (`it's...complicated`). Reads as performance.
33
34## Banned Sentence Openers
35
36- "Importantly," "Notably," "Crucially," "Interestingly"
37- "Furthermore," "Moreover," "Additionally"
38- "It's worth noting that…" / "It bears mentioning…"
39- "In essence," "Ultimately," "At its core"
40- "In conclusion," "To summarize," "In summary"
41- "Great question!" or any sycophantic acknowledgement
42
43If the sentence after the opener stands on its own, delete the opener. It almost always does.
44
45## Banned Vocabulary
46
47These words appear far more often in machine-generated prose than human prose. Each has a fix-direction.
48
49| Word / Phrase | Why it's a tell | Fix direction |
50|---|---|---|
51| `delve` | Diagnostic on its own | use `look at`, `examine`, or just delete |
52| `leverage` (verb) | Corporate-AI fingerprint | use `use` |
53| `robust`, `comprehensive`, `seamless`, `streamlined` | Empty positive adjectives | name the specific property |
54| `cutting-edge`, `innovative`, `state-of-the-art` | Marketing puff | describe what it actually does |
55| `elegantly`, `seamlessly`, `gracefully` | Empty positive adverbs | delete the adverb |
56| `modulo` (in non-math prose) | Claude tic | use `except for` or `aside from` |
57| `stakeholders` | Generic abstraction | name the people or roles |
58| `passionate about`, `excited to`, `thrilled to` | Performative enthusiasm | state what was done |
59| `it's worth noting` | Throat-clearing | delete; state the thing |
60| `at the end of the day` | Verbal filler | delete |
61| `move the needle`, `circle back`, `align on` | Corporate cliche | name the actual action |
62
63## Banned Structural Reflexes (Surface Layer)
64
65These are mechanical enough to catch on a scan; deeper rhythm/structure tells live in `ai-tells-review`.
66
67- **Tricolon reflex.** Three-item lists by default ("fast, reliable, and scalable"). If two items would do, use two. If five, use five.
68- **"This is not X — it's Y" contrastive bold.** The mirrored "Why not pure JSON: / Why not pure scriptblocks:" pattern. Delete the contrast or merge into a single sentence.
69- **Decision-stamp formality.** `**Decided 2026-05-18:**` style headers in personal notes. Real humans rarely formalize their own decisions like a court ruling.
70- **Self-narration.** Sentences that comment on what the doc is doing: "That's the dashboard's reason for existing." / "(this is where we start)". Delete and let the doc do the thing instead of describing itself.
71- **Compliment sandwich.** Soften → criticize → soften. State the criticism directly.
72- **Three consecutive sentences saying the same thing in different words.** Delete two of them.
73
74## Banned Visual Reflexes
75
76- **ASCII art boxes for layered diagrams.** When the diagram is more than two boxes, use mermaid or just describe the layers in prose.
77- **Italicized rhetorical emphasis.** *missing*, *separately*, *on this dashboard*. Italics are for terms-of-art and titles, not tone. One per long doc, max.
78- **Bold for emphasis on every paragraph.** Bold should mark scannable structure (headers, key terms), not feeling.
79
80## Coined-Phrase Tic Check
81
82Before submitting, scan for any minted phrase repeated 3+ times across the doc. Examples seen in real artifacts: "proves the framework," "earns its place," "5-second glance." Humans usually rephrase the second or third use; models don't. Rephrase all but the first occurrence.
83
84---
85
86## What This Skill Does Not Do
87
88- Catch rhythm patterns (uniform sentence length, paragraph-level parallelism, gerund cascade)
89- Catch hedging cadence and manufactured both-sidesing
90- Catch aphoristic closers and summary-of-summary endings
91- Replace voice skills like `nick-voice` — this subtracts machine fingerprints, it doesn't add personality
92
93For all of the above, use `ai-tells-review`.
94
95## Output Shape
96
97The scan is invisible by default — it runs before the prose is presented and the prose comes out clean. If a tell can't be cleanly fixed (e.g. the contrastive-bold pattern is load-bearing for the argument), flag it inline:
98
99> NOTE: kept "X — Y" construction at line 42; the contrast carries the argument and rewriting flattens it.
100
101This gives the user one place to check whether to push back on a specific judgment.
nickcrew