# Ai Tells > Reviews any draft for AI writing tells and produces a clean rewrite. Catches vocabulary fingerprints (delve, tapestry, leverage, robust, multifaceted, navigate, foster), structural fingerprints (negative parallelism like "It's not X, it's Y", bullet-everything, Title Case headings, adjective triads), tone fingerprints ("Great question!", "I'd be happy to help", "In conclusion"), and punctuation f… - **Type:** Skill - **Install:** `agentstack add skill-kdgbalmer-ai-tells-ai-tells` - **Verified:** Yes — security-reviewed for prompt injection and unsafe behavior - **Seller:** [kdgbalmer](https://agentstack.voostack.com/s/kdgbalmer) - **Installs:** 0 - **Category:** [Agent Skills](https://agentstack.voostack.com/c/agent-skills) - **Latest version:** 0.1.0 - **License:** MIT - **Upstream author:** [kdgbalmer](https://github.com/kdgbalmer) - **Source:** https://github.com/kdgbalmer/ai-tells ## Install ```sh agentstack add skill-kdgbalmer-ai-tells-ai-tells ``` Requires the [AgentStack CLI](https://agentstack.voostack.com/docs/cli). Works with Claude Code, Cursor, and any MCP-compatible agent. ## About # 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:Signs_of_AI_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](https://github.com/kdgbalmer) - **Source:** [kdgbalmer/ai-tells](https://github.com/kdgbalmer/ai-tells) - **License:** MIT Install and usage instructions live in the source repository linked above. ## Pricing - **Free** — Free ## Security capabilities Automated source analysis of v0.1.0 — what this tool can access: - **Network access:** no - **Filesystem access:** no - **Shell / process execution:** no - **Environment & secrets:** no - **Dynamic code execution:** no *"Yes" means the capability is present in the source — more access means more to trust, not that it is unsafe.* ## Versions - **0.1.0** — security scan: passed — Imported from the upstream source. ## Links - Listing page: https://agentstack.voostack.com/l/skill-kdgbalmer-ai-tells-ai-tells - Seller: https://agentstack.voostack.com/s/kdgbalmer - Browse the marketplace: https://agentstack.voostack.com/browse --- Listed on AgentStack — the marketplace for AI agent skills and MCP servers. Every listing is security-reviewed. Creators keep 70%.