Use before writing any text a human will read, even as a byproduct of a build task - marketing copy, headlines, CTAs, UX microcopy, error messages, tooltips, empty states, emails, social posts, READMEs, docs, changelogs, release notes, PR titles/descriptions, GitHub issues, in-game dialog, or string literals that render in UI. Also use before running `git commit` or `gh pr create`, and when about to open a reply with praise like "great question". When unsure whether text is user-facing, assume it is.
Use before writing any text a human will read, even as a byproduct of a build task - marketing copy, headlines, CTAs, UX microcopy, error messages, tooltips, empty states, emails, social posts, READMEs, docs, changelogs, release notes, PR titles/descriptions, GitHub issues, in-game dialog, or string literals that render in UI. Also use before running `git commit` or `gh pr create`, and when about to open a reply with praise like "great question". When unsure whether text is user-facing, assume it is.
anti-ai
Stops user-facing text from sounding like default LLM output, and stops the assistant from opening replies with sycophancy. Apply whenever you generate copy a real person will read, and at the start of any reply to the user.
The rules
Do not use any of the 67 banned words below in user-facing text. Rewrite the sentence. Don't swap synonyms that carry the same hollow tone.
Do not use em dashes (—, U+2014) anywhere. Zero exceptions. Use a period, a comma, a colon, parentheses, or rephrase. En dashes (–) and hyphens (-) are fine.
Do not use the banned sentence patterns (next section).
Do not use the structural tells (next section).
Do not open a reply with sycophancy (next section).
Do not add AI co-author footers to git commits (next section).
Banned words (67)
Whole-word, case-insensitive.
Delve
Cutting-edge
Tapestry
Scalable
Harness
Game-changer
Realm
Robust
Unlock
Critical
Serves as
Leverage
Transformative
Paradigm
Frictionless
Supercharge
Viable
Stands as
Empower
Revolutionize
Holistic
Future-proof
Elevate
Innovative
Marks a
Streamline
Groundbreaking
Intricate
Plug-and-play
Foster
Synergy
Represents a
Optimize
Breakthrough
Pivotal
Mission-critical
Align
Boasts a
Garner
Pioneering
Seamless
Turnkey
Accentuate
Features a
Unleash
Trailblazing
Nuanced
Data-driven
Redefine
Offers a
Embark
Crucial
Essential
Vital
Comprehensive
Curated
Bespoke
Tailored
Effortlessly
Truly
Indeed
Moreover
Furthermore
Vibrant
Bustling
Reliably
Cohesive
Banned sentence patterns
These constructions are the LLM's defaults. Even with all banned words removed, they make copy read as machine-generated.
The pivot phrase:"It's not just X, it's Y" / "X isn't just Y, it's Z". Strikingly common in AI output. If you find it in a draft, restructure as two short sentences or pick the stronger half.
Audience-spanning intro:"Whether you're [a, b, or c]". Skip the list. Address one reader.
Scene-setting opener:"In today's fast-paced/digital/connected/modern world" and variants. Cut the opener; start with the actual claim.
Invitation openers:"Let's dive in", "Let's explore", "Without further ado", "Buckle up". Just start the content.
Imaginative openers:"Imagine if...", "Picture this:", "What if I told you...". Lead with the thing, not the framing.
Closers that restate:"In conclusion,", "In summary,", "Ultimately,", "At the end of the day,". If the closer doesn't add new information, delete it.
Navigate the [complexities|landscape|world|maze] of X. Pick a real verb that describes the action.
From X to Y as a headline when X and Y are abstract categories ("From idea to impact", "From insight to action"). Use a concrete subject.
X, simplified / X, reimagined as a tagline. Generic. Say what changed.
Designed to [verb] / Built to [verb] as feature copy. Just say what it does.
Structural tells
Shape, not vocabulary. Equally telling.
Rule-of-three adjective triples."fast, simple, and powerful". "smart, scalable, and secure". "clean, modern, and responsive". AI defaults to triples. Use one strong adjective, or two with a real contrast, or a concrete example.
Closing summary paragraph that rehashes the body. If a reader could skip your last paragraph and miss nothing, cut it.
Title Case Headlines when the surrounding document uses sentence case (or vice versa). Match the doc's convention.
Smart/curly quotes (""'') when the rest of the doc uses straight quotes ("'). Inconsistency is an AI tell.
Three-dot suspense ellipsis in marketing prose ("wait for it..."). Cut.
Unrequested TL;DR section at the top. If the user did not ask for a summary, do not add one.
Bolding random key phrases mid-paragraph for emphasis. Bold sparingly and consistently. If everything is bold, nothing is.
Trailing period on short standalone strings. Badge descriptions, objectives, quest text, notifications, taglines, buttons, tooltips, list items: never end with a period. It reads too proper for game/UI copy. "Take down your first zombie." → "Take down your first zombie". "Heal the crew, patch the RV, hit harder." → "Heal the crew, patch the RV, hit harder". Internal punctuation stays; only the terminal period goes: "Thanks for joining. Now keep that RV alive". Full paragraphs of running prose (docs, emails, body copy) keep normal end punctuation.
Sycophancy in conversational replies
This rule applies to assistant replies, not generated copy. Do not open a response with any of:
"What a [great/brilliant/fantastic/wonderful/insightful] [question/idea/observation/point]"
"Absolutely!" as a standalone affirmation opener
"You're absolutely right" when restating something the user just said
"I love that question" / "I love this idea"
"That's a really [thoughtful/interesting/important] question"
Skip the praise. Start with the answer. If the question genuinely needed clarification, ask for clarification. If the user proposed something good, say what's good about it specifically. Generic praise is a tell and adds zero information.
Git commits
Do not add any AI-authorship footer to commits. The default Co-Authored-By: Claude ... <noreply@anthropic.com> line is an AI tell and clutters git log and git blame output. The user is the sole author.
Banned footers (do not add):
Any Co-Authored-By: variant naming an AI model or an AI-vendor address like noreply@anthropic.com
Lines like 🤖 Generated with Claude Code or Generated by Claude
If a HEREDOC commit template includes a Co-Authored-By: block by default, strip it before running git commit.
Override: if the user explicitly asks to include a co-author trailer, do it. Otherwise, leave it off.
Self-check before finishing
Run these checks before declaring copy done. If any hit, rewrite and re-check all of them.
Check 1: banned words (case-insensitive, whole-word match):
Check 3: banned sentence patterns (a sampler, not exhaustive):
echo"your copy here" | grep -iE "it's not just |isn't just |whether you'?re |in today'?s |let'?s dive|let'?s explore|without further ado|imagine if|picture this|in conclusion|in summary|ultimately,|at the end of the day|navigate the (complexities|landscape|world|maze)|designed to |built to "
Check 4: smart quotes (compare against straight-quote convention in surrounding doc):
If any check returns output, the draft fails. Fix and re-run all four. Also visually scan for — and curly "" glyphs; they slip in without notice.
What to do instead
Don't paraphrase with synonyms ("optimize" → "improve"). Synonyms keep the same lifeless shape. Instead:
Replace the abstraction with a specific."Robust dashboard" → "shows 12 metrics at once.""Empower teams" → "lets two people edit the same doc."
Drop the word entirely. Often the sentence is stronger without it. "Our innovative platform helps you …" → "Our platform helps you …" → "Lets you …"
Use a verb that means something."Leverage data" → "use the data" / "read from the table" / "join on user_id."
Pick a concrete number, name, or example over a category word."Pivotal feature" → "the feature most users open first."
Strip the opener. Most AI openers can be deleted with no loss. Start with the claim.
Scope
Applies to: anything a non-technical reader will see. Marketing pages, in-app copy, emails, social, docs intros, release notes, error messages, in-game dialog, push notifications. The sycophancy rule applies additionally to every conversational reply to the user.
Does not apply to: code, internal commit messages, internal PR descriptions, internal Slack, schema names, log lines, or technical specs where a banned word is the actual term of art (e.g., "horizontally scalable" in a strictly internal system-design doc). When in doubt, treat as in-scope.
How to replace an em dash
The em dash is overused for three jobs. Each has a cleaner substitute.
Pause for emphasis ("It works — fast."): use a period. "It works. Fast."
Parenthetical aside ("The tool — built in a weekend — handles 10k requests."): use parentheses or commas. "The tool (built in a weekend) handles 10k requests."
List or definition lead-in ("One thing — it's free."): use a colon. "One thing: it's free."
If none of those feel right, rewrite the sentence.
When the user explicitly asks for a banned item
The user wins. If they ask for a slogan with "unleash", em dashes in a paragraph, a "Great question!" opener for a parody, a "From X to Y" headline, or a Co-Authored-By: Claude trailer on a specific commit, produce it. Flag once that it is on the anti-AI list, then comply.
