AI tell killer.
A checklist of the specific patterns that flag writing as AI in 2026. Pick the ones to kill. Get a paste-in prompt for ChatGPT, Claude, or Gemini. Also gives you a Claude Code rules block and a Node.js lint script for catching anything that slips through.
deeper dive on em dashes: the em dashes guide
· pick the tells to kill
7 of 20 selected
· Word choice
· Sentence and section structure
· Markdown and formatting
· Reply behavior
· your paste-in
Apply these rules to every reply you write for me. They aim to remove the specific patterns that flag content as AI-written.
1. Don't use em dashes (U+2014, the long one) or en dashes (U+2013, the shorter one) in prose. In code that legitimately needs these characters (regex patterns, Unicode escapes, test fixtures), preserve them. For replacements: two clauses become a period and a new sentence, parentheticals use commas or parens, list intros use a colon, compound modifiers use a hyphen, numeric ranges use "to" or a hyphen.
Self-review: Scan output for U+2014 and U+2013. If found, rewrite that sentence from scratch. Do not character-swap.
2. Don't use the word "comprehensive". If completeness needs to be conveyed, say what is actually covered ("covers X, Y, and Z") or use words like "full", "all", or just remove the adjective.
Self-review: Scan for "comprehensive". Rewrite the surrounding sentence rather than swapping the word.
3. Avoid the following words unless precisely required: "delve", "tapestry", "multifaceted", "leverage" (as a verb), "unlock" (as a metaphor), "navigate" (as a metaphor), "embark", "robust", "elevate", "transformative", "pivotal", "testament", "intricate", "underscore", "harness", "showcase", "boasts". Prefer plain alternatives: explore -> look at, leverage -> use, unlock -> open or enable, navigate -> work through, robust -> sturdy or reliable, transformative -> changes how X works.
Self-review: Before replying, scan for any of the banned words. Rewrite the sentence to avoid them.
4. Don't open paragraphs with: "It's worth noting", "It's important to note", "As someone who", "In today's digital age", "In the world of", "Ever wondered why", "When it comes to". Start with the actual claim or observation. If a hedge is needed, place it later in the sentence.
Self-review: Check the first sentence of each paragraph. If it matches one of the patterns above, rewrite it.
5. Avoid "Moreover", "Furthermore", "Additionally", "In conclusion", and "Ultimately" as transition openers. Either connect ideas directly without a transition word, or use a plain "Also" or "And" if a connector is needed.
Self-review: Scan paragraph and sentence beginnings for the listed transitions. Remove or restructure.
6. Don't open replies with praise of the question or the user. No "great question", "excellent point", "that's a fascinating thought", "you're absolutely right". If the user is wrong, push back. If the user is right, just confirm and continue. If the user's question is unclear, ask for clarification rather than guessing and praising.
Self-review: Check the first sentence of replies. If it praises the user or the question, delete it.
7. Don't fabricate specifics. If you don't know a number, a name, a citation, or a date, say so explicitly: "I don't have a reliable source for this number" or "I'm not certain of the exact date." Don't fill gaps with plausible-sounding inventions. If a citation is required and you don't have one, say none is available.
Self-review: For any specific number, date, name, or citation in the output, ask: am I sure of this, or am I extrapolating? If extrapolating, replace with an honest uncertainty statement.
Before sending any reply, run through the self-reviews above. If any rule was violated, rewrite the affected sentence from scratch rather than character-swapping.What an AI tell is, and why removing them matters
An AI tell is a writing pattern that LLMs produce at much higher rates than human writers. Em dashes in casual prose. The word "comprehensive" on every other page. Mechanical bullet lists of exactly three items. "Great question!" openers that praise the user before answering. None of these are wrong on their own, but they cluster in AI output at densities that pattern- matching readers detect quickly.
In 2026, that pool of pattern-matching readers is growing. Editors, hiring managers, professors, recruiters, and a meaningful slice of the general public now read suspiciously and look for the tells. For prose meant to feel personal (LinkedIn posts, marketing pages, cover letters, casual emails, op-eds) the cost of leaving the tells in is the AI-author label. The cost of removing them is small.
The three output formats
The same rule set powers three different surfaces. Use whichever fits the workflow.
- Prompt: a numbered paste-in for the custom-instructions field of Claude, ChatGPT, or Gemini. Each rule has its own line plus a self- review step the model runs before replying. This is the 90- second setup.
- Claude Code rules: a markdown block formatted for
~/.claude/CLAUDE.md. Carries the rules across every Claude Code session on every project. Includes code-safe exceptions so it won't break regex patterns, Unicode escapes, or test fixtures. - Lint script: a self-contained Node.js script that walks a file or directory and flags any line matching the regex patterns for the tells you picked. Exits non-zero on findings. Wire into a pre-commit hook or CI for prose-heavy repos. Not every tell has a regex form (sycophancy and rule-of-three lists can't be detected by pattern), so the script covers the subset that can.
How to use this with each LLM
- Claude (web / desktop): Settings, Profile, Custom Instructions. Paste the prompt. Save. The rules apply to every chat on that account from now on.
- ChatGPT: Settings, Personalization, Customize ChatGPT, "How would you like ChatGPT to respond?" field. Paste. Save.
- Gemini: Saved Info, or create a default Gem with the prompt as its system instruction. Gemini is less reliable on long outputs; pair with the lint script.
- Claude Code: save the Claude Code rules block to
~/.claude/CLAUDE.md. Applies to every project and every session.
Single-tell paste-in pages
If you only want to kill one specific behavior, each of these pages gives you the paste-in for that one rule. The full tool above covers all 21 in one paste; these are for when you searched for a specific fix.
- Stop ChatGPT from being a yes man (kill sycophancy).
- Stop ChatGPT from hallucinating (force honest uncertainty).
- Stop ChatGPT from restating your question (kill the preamble).
- Stop ChatGPT from using bullet points (force prose paragraphs).
- Stop ChatGPT from saying "comprehensive" (the 2026 standout word tell).
FAQ
What is an "AI tell"?
A pattern in writing that flags content as AI-written to readers paying attention. Em dashes overused in casual prose, the word "comprehensive" appearing where any other word would do, sycophantic openers like "Great question!", rule-of-three lists, mechanical bold on every key term. None of these are wrong on their own, but at high enough density they let a reader pattern-match the output to LLM origin. In 2026 the number of readers doing that pattern-matching is growing fast, so removing the strongest tells is a cheap way to make content read more natural.
Do these paste-in prompts actually work? I thought LLMs ignore negative instructions.
Standalone "never do X" instructions are weak. The prompts this tool generates combine three things per rule: an explicit ban, a specific replacement or alternative, and a self-review step that asks the model to scan its own output before replying. The combination catches most cases. The lint script catches anything that slips through. For tells with detectable patterns (em dashes, the word "comprehensive", curly quotes, certain transitions) the lint script is the last-mile safety net.
Which LLM follows these rules most reliably?
In current 2026 testing, Claude 4.7 and ChatGPT GPT-5.5 follow the rules reliably when the prompt sits in custom instructions or a system prompt (not just a user turn). Gemini 3.1 Pro is somewhat less reliable on long outputs, where it tends to revert in the second half of a response. For Gemini specifically, run the lint script over the output before shipping.
Why is the word "comprehensive" on the list? It's just a word.
In 2026, "comprehensive" appears in LLM output at a rate that human writers basically never match. Once you notice it, you stop seeing genuine uses and start seeing the AI tic. The fix is rarely word-substitution; usually the right rewrite is to say what the content actually covers ("covers cost, latency, and reliability") rather than asserting it is "comprehensive." Same goes for "robust", "transformative", "leverage", and the rest of the vocab cluster.
Are em dashes still a useful tell? I heard GPT-5.5 suppresses them.
GPT-5.5 does suppress em dashes more aggressively than earlier models, which weakens em dashes as a sole signal. But em dashes still appear in older Claude / Gemini output, in content from older models, and in fine-tuned models that re-introduce them. Em dashes are also a strong reader signal even if model output volume drops; many readers in 2026 use them as a first-pass detector. Keep the rule on. For the full reasoning, see the dedicated em dashes guide.
Won't this make my writing sound worse? Some of these patterns are legitimate.
They are. The list isn't "these are bad in all writing"; it's "these are AI-signature when used at high density." A human writer who uses one em dash every 500 words reads as a confident stylist. An LLM that uses four em dashes per paragraph reads as a model. The rules tell the LLM not to default to the pattern. For content that genuinely needs the pattern (a formal essay where em dashes are appropriate, technical docs where bullet lists are correct), drop the relevant rule before generating.
What about code? Won't a strict 'no em dashes' rule break my regex patterns?
No. The generated Claude Code rules block includes explicit code-safe exceptions: regex patterns that literally match the banned characters, Unicode escape sequences (`\u2014` is fine; only the rendered character is banned), test fixtures verifying behavior with the banned input, and data files where preserving the original is required for correctness. The rules apply to prose, not to code that load-bears on the character.
Where does my checklist data go?
Nowhere outside your browser. The tool runs entirely client-side. Your selections persist in LocalStorage so they're there next time you open the page. No network request is made when you toggle tells, switch output modes, or copy/download. No analytics event includes which tells you picked.
Companion tool: the AI Output Linter takes the diagnostic side. Paste AI text into it and see every tell highlighted with line numbers, so you can verify what slipped through after the paste-in prompt did its job. Same tells library, different side of the workflow.
More on this topic: the em dashes guide covers the single most-checked tell in detail. The system prompt primer explains where each LLM's custom-instruction field lives. For cost-per-prompt math across models, the AI cost calculator runs the numbers.