Ai Tells

AI tells

A second-pass checker that flags AI writing tells in a draft and produces a clean rewrite. The catalogue is compressed from Wikipedia's "Signs of AI writing" page plus the goblin-era 2026 update.

Modes

Two modes:

1. Audit. Flag every tell, name the category, suggest a replacement. No rewrite.
2. Rewrite. Produce a clean version with all tells removed, preserving meaning, intent, and voice.

Default to audit + rewrite unless the user asks for one mode.

The catalogue

Walk every category below. For each match in the input, note the exact phrase, the category, and a proposed replacement.

Vocabulary tells

These words appear at AI-anomalous frequency in machine-generated text. Replace with plainer alternatives unless context absolutely demands them.

Tier 1, high-confidence AI markers: delve, tapestry, multifaceted, pivotal, intricate, robust, vibrant, meticulous, nuanced, leverage, foster, navigate, underscore, showcase, ensure, realm, garner, bolster, enduring, elevate, unwavering, testament, journey, landscape, ecosystem, paradigm.

Tier 2, context-dependent: crucial, key, vital, significant, essential, comprehensive, holistic, seamless, dynamic, innovative, transformative, cutting-edge, state-of-the-art, harness, embrace, embark, dive into, dive deep.

Tier 3, promotional language (cut especially in marketing or about-page contexts): boasts a, nestled in, in the heart of, renowned for, exemplifies, stands as a testament, serves as a reminder, represents a shift, marks a turning point, indelible mark, deeply rooted, rich, profound, enhancing, showcasing, commitment to excellence.

When one of these appears, ask: is there a plainer word that means the same thing? If yes, swap it.

Structural tells

Negative parallelism ("not X, it's Y") is the most overused AI structure. Cut every instance unless it's a direct quote from a real human source. Examples:

• "It's not [a tool], it's [a thought partner]"
• "You're not [an X], you're [a Y]"
• "Not just A, but B"
• "It's not about A. It's about B."

Bullet-everything. A 1-3 sentence answer broken into bullets, or a markdown header used for a paragraph-length response. Convert to flowing prose.

Rule of three. Adjective triads ("creative, thoughtful, and deeply considered") used for false comprehensiveness. Cut to one or two specific adjectives.

Outline-style conclusions. Final paragraphs that read "Despite [challenges], [subject] continues to [vague positive outlook]." Replace with a sharp specific kicker, or no conclusion at all.

Title Case headings when sentence case is the convention. Convert.

Skipping heading levels (H2 directly to H4 with no H3). Fix the hierarchy.

Inline-header lists (- **Term:** description repeated mechanically). Use only for genuine definition lists. Otherwise convert to prose.

Excessive boldface. Bolding every key term mechanically. Cap bold at roughly two instances per section, only on phrases the reader needs to find by skimming.

Tone and opener tells

Cut these phrases and the sentences they live in:

• "Great question!"
• "Absolutely!"
• "Certainly!"
• "Of course!"
• "I'd be happy to help."
• "I'd love to help."
• "I'm here to help."
• "It's important to note that…"
• "It's worth noting that…"
• "It should be noted that…"
• "Interestingly,"
• "In conclusion,"
• "In summary,"
• "To summarize,"
• "To wrap up,"
• "Furthermore,"
• "Moreover,"
• "Additionally,"
• "However," used as the opening word of multiple paragraphs.

Test: if the sentence still works without the opener phrase, delete the phrase. If the sentence collapses without it, the sentence is filler, so delete the whole sentence.

Punctuation tells

Em dash overuse. More than 2-3 em dashes per 500 words, or em dashes appearing at every clause boundary. Replace most with commas, full stops, or hyphens with spaces. Genuine asides can keep theirs.

Curly quotes in plain-text or markdown contexts where the surrounding ecosystem uses straight quotes. Convert.

Knowledge-cutoff disclaimers. "As of my knowledge cutoff…", "I may not have current information…". Cut from final output.

Sycophantic openers. "What a wonderful question!", "I love this prompt!". Cut.

Content and analysis tells

Vague attribution. "Many experts say…", "Studies have shown…", "It is widely believed that…". Cite a specific source or delete the claim.

Superficial -ing analysis. Clauses that add no information: "highlighting their significance", "emphasizing their role", "underscoring their importance". Cut the whole clause.

Hedge stacking. "May potentially possibly suggest." Pick the strongest verb that's still accurate.

Filler hedges. "Somewhat", "relatively", "arguably", "perhaps", "potentially". Cut where the claim still holds. If the claim doesn't hold without the hedge, the claim is too weak. Sharpen or delete.

Era-specific tells (current as of May 2026)

Active right now:

• Goblin, gremlin, raccoon, troll, ogre, pigeon. OpenAI hard-banned these in May 2025; if they appear, they leaked through.
• "It's not X, it's Y." Still rampant.
• Markdown-bullet-everything. Still rampant.
• "I'd be happy to help" openers. Still rampant.
• Title Case section headings. Still rampant.

Mostly trained out:

• "Delve". Paul Graham's 2024 viral tweet pushed OpenAI to train it down. Now uncommon.
• Em dashes at every clause boundary. OpenAI shipped a fix in November 2025. Less common but still leak through.
• "Tapestry". Peak 2023. Now rare.

Update this section quarterly against the Wikipedia source. Demote trained-out tells to the second list, but never delete from the catalogue. Tells are cyclical and may return.

Workflow

1. Read the input fully. Don't skim. Density and combination matter more than any single tell.
2. Pass 1, flag. Walk every category. For each match, note the phrase, the category, and a proposed replacement.
3. Pass 2, score. Estimate tells per 100 words. Above 3 is heavy. Above 6, recommend a rewrite from scratch rather than editing.
4. Pass 3, rewrite. Maintain the writer's intent and voice. Don't substitute one set of AI tells for another.
5. Pass 4, self-audit. Read the rewrite. Ask: what still feels AI? Fix it. Repeat until clean.

Output format

Return three blocks:

1. Audit. A bulleted list of every tell flagged. Format: [category] "exact phrase" → "proposed replacement".

2. Density score. "X tells per 100 words. Light / Moderate / Heavy / Unsalvageable."

3. Clean rewrite. The final version with all tells removed.

If the input scores Unsalvageable, skip the rewrite and tell the user the draft needs to be redone from a fresh outline rather than edited.

What this skill does not do

• It does not detect whether a text was AI-generated. AI detectors are unreliable. This skill assumes the input may be AI-assisted and cleans it regardless.
• It does not substitute for human editing. It catches surface tells. It cannot detect hollow thinking or missing voice. Those need a human pass.
• It does not produce style. It produces absence of AI style. Pair it with a voice guide for the affirmative side.

Source and credit

• Wikipedia: Signs of AI writing. https://en.wikipedia.org/wiki/Wikipedia:SignsofAI_writing

Built by Kyle Balmer at AI with Kyle. MIT licensed. Last updated 2026-05-04. Free to use, share, and modify. If the list dates, update it.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: kdgbalmer
• Source: kdgbalmer/ai-tells
• License: MIT

Install and usage instructions live in the source repository linked above.