| name | ai-writing-tells |
| description | Scan prose for AI writing tells, report numbered findings by category with line references, wait for user approval before editing, and rescan after approved fixes. |
AI Writing Tells
Use this skill when the user asks to scan prose for AI writing tells, chatbot-style wording, generic polish, or minimal "deslop" editing in docs, prompts, comments, commit text, or other human-facing writing.
Scope
- Audit the specified file, selected text, or the file the user is actively editing.
- Prefer exact line references and short quoted phrases.
- Do not edit until the user explicitly confirms which items to address.
Workflow
- Determine the target text.
- If the user named a file or pasted text, use that.
- If the user refers to "this file" or "current file", infer the active file only when it is clear from context.
- If the target is ambiguous, ask one concise question before scanning.
- Read only the target text.
- Scan in this order: Vocabulary, Patterns, Formatting.
- Use judgment. Flag only wording that feels generic, padded, promotional, or mechanically polished in context.
- Present findings in the required output format below.
- Ask:
Which items should I fix? (all / specific numbers / none)
- If the user approves fixes, edit only the approved items.
- Re-scan after edits and report the updated findings and score.
- If the updated score is below 8, ask:
Another round? (yes / no)
- Stop when the user says stop or the score reaches 8 or 9. Do not chase a 10 unless the user explicitly wants a voice rewrite.
What To Scan For
Vocabulary
Flag any of these when they read like filler rather than precise wording:
additionally (especially at sentence start)
align with
boasts (meaning "has")
bolstered
crucial
comprehensive
cutting-edge
delve
emphasizing
enduring
enhance
encompassing
features (meaning "has")
foster / fostering
garner
groundbreaking
highlight (as a verb)
innovative
interplay
intricate / intricacies
key (as an adjective)
landscape (abstract noun)
maintains (meaning "has")
meticulous / meticulously
nuanced
offers (meaning "has" or "gives")
pivotal
profound
renowned
robust
seamless / seamlessly
showcase / showcasing
tapestry (abstract noun)
testament
transformative
underscore (as a verb)
valuable
vibrant
vital
Note: AI vocabulary shifts over time. Delve peaked earlier; later drafts leaned harder on words like enhance, highlighting, and showcasing. Flag words that feel like filler in context, not words used precisely.
Patterns
- Significance puffery:
stands as, serves as, is a testament to, plays a vital role, underscores its importance, reflects broader, setting the stage for, marking the, shaping the, evolving landscape, indelible mark, deeply rooted
- Superficial
-ing analysis: sentences that end in dangling present participles adding vague commentary, such as highlighting its importance, ensuring quality, reflecting the broader trend, contributing to the field, fostering innovation
- Negative parallelisms:
not just X, but also Y, not only ... but ..., it's not ... it's ...
- Rule of three: formulaic three-item lists where two would do
- Copula avoidance:
serves as instead of is, boasts instead of has, represents instead of is, stands as instead of is, features instead of has, maintains instead of has, offers instead of gives
- Em dash overuse: too many em dashes where commas, parentheses, or periods would read more naturally
- Elegant variation: cycling through fancy synonyms just to avoid repeating a direct word
- Promotional tone:
nestled in, in the heart of, diverse array, rich heritage, natural beauty, commitment to
- Vague attribution:
experts argue, observers note, industry reports suggest, observers cite
- Despite-challenges formula:
Despite ... faces challenges ... followed by an optimistic recovery
- Didactic disclaimers:
it's important to note, worth noting, it's crucial to remember
- Formulaic headings:
Challenges and Future Prospects, Legacy and Impact, Significance and Influence
Formatting
- Curly quotes or apostrophes in contexts that expect straight quotes
- Markdown syntax leaking into non-Markdown text, such as stray
**bold** or [links](url)
- Excessive boldface used only for emphasis
- Inline-header vertical lists where every item starts with a bold label and colon even though a plain list would be clearer
Required Output
- Group findings by category.
- Use continuous numbering across categories.
- Include line numbers and a short quote or phrase for each finding.
- Omit empty categories if there are no findings there.
- Always include
Summary and Human Writing Score.
## AI Writing Tells Found
### Vocabulary
1. Line X: "delve" - consider a simpler alternative
2. Line X: "fostering" - filler verb
### Patterns
3. Line X: Negative parallelism - "not just X, but Y"
4. Line X: Significance puffery - "serves as a testament"
### Formatting
5. Line X: Curly quotes in a code context
### Summary
X items found across Y categories.
### Human Writing Score: X/10
Then ask: Which items should I fix? (all / specific numbers / none)
Scoring Guide
10: Reads like a specific human wrote it and has a clear voice.
9: Clean. No real tells.
8: One or two minor patterns, but nothing most readers would notice.
7: Competent but generic. A trained eye would catch several tells.
6: Obviously polished by AI. Multiple vocabulary or pattern flags.
5 or below: Reads like a chatbot wrote it.
Do not inflate scores. A score of 10 is not the goal.
Fix Rules
- Replace flagged words with simpler, more direct alternatives.
- Cut superficial
-ing phrases instead of rewriting them into new filler.
- Flatten negative parallelisms into direct statements.
- Break rule-of-three lists if the third item adds nothing.
- Replace
serves as or stands as with is where appropriate.
- Change only the approved items. Do not refactor surrounding text.
Rescan Rules
- After fixing, re-scan the same text and report the updated findings and score.
- If the score is below 8, ask:
Another round? (yes / no)
- Do not auto-fix another round without confirmation.
Tell list adapted from Wikipedia: Signs of AI writing, licensed under CC BY-SA 4.0. Modified and extended for use as a code editor skill.
This file is licensed under CC BY-SA 4.0.