ai-copywriter: A Claude Code Skill for Marketing Copy That Doesn't Sound Like AI

Ask most models to write a headline and the result is technically fine and instantly forgettable: a rounded, hedge-everything sentence that reads like marketing copy about marketing copy. ai-copywriter, a claude code skill for marketing copy built by developer Mickey Haslavsky, is an attempt to fix that at the source rather than editing around it afterward. It bundles an actual copywriting method with a detailed detector for AI-sounding prose, and applies both to the copy it writes itself before handing it back.

What this claude code skill for marketing copy actually does

The skill does not name classic formulas like AIDA or PAS. Instead it opens every job with two questions, in order: what is the reader feeling at the exact moment this line reaches them, and what is the simplest way to explain the thing being sold. Before drafting, it runs an intake, asking who the reader specifically is (not "founders," but a named situation and what they'd type into a search box at 11pm), what category the product gets filed under in the reader's head, and what the actual story behind the copy is, since a real story with a specific number in it is the raw material a fabricated one can't fake.

That intake feeds format-specific rules for headlines, short descriptions, microcopy (buttons, error messages, empty states), subject lines, and LinkedIn posts. The rules are concrete rather than aspirational: buttons name the result of the action ("Send invoice," not "Submit"), error messages say what broke and how to fix it without blaming the reader, and headline output comes back as five to ten variants across different angles with a one-line note on which to ship and why. There's also a longer strategic blog post template in the repo's references/ folder, aimed at founder-voice posts rather than short-form copy.

One rule runs under everything: no invented product facts. Any number, name, or claim in the output has to trace back to something the user supplied or the source material, not something that sounds plausible.

The humanizing half, mechanically

"Human tone" here isn't a vibe setting. It's a list of 33 specific, checkable patterns pulled from Wikipedia's "Signs of AI writing" guide, the same reference WikiProject AI Cleanup editors use to flag machine-written Wikipedia edits. The skill runs every draft, including its own, through that list before handing it back.

Concretely, that means catching promotional inflation ("boasts," "showcases," swapped for plainer verbs like "is" or "has"), stock AI vocabulary ("actually," "additionally," "landscape"), rule-of-three list padding, manufactured aphorisms, vague attributions, negative-parallelism constructions ("it's not just X, it's Y"), and overused em dashes. It also flags title-case headers and stacked rhetorical openers like "Honestly? It depends." The fix isn't a find-and-replace pass. The instructions tell the model to preserve every fact in the original while rewriting the shape entirely, compressing the dull parts and dwelling where a person actually would.

There's a real exception worth noting: if a user supplies a sample of their own past writing, the skill matches that sample's habits instead of scrubbing every AI tell by default. If the sample uses em dashes at a certain rate, the rewrite keeps that rate. Matching a real person's voice outranks the general style rules.

How this differs from x-skills

AIOS Guide already covers x-skills, a Claude Code skill bundle with its own Humanizer sub-skill for stripping AI writing tells out of tweets and threads. The overlap is real: both attack the "obviously AI-written" problem head-on, and both ban em dashes and stock AI vocabulary as a baseline rule.

The difference is scope. x-skills is X-specific: nine skills built around one platform's constraints, tuned to a 280-character hook and the feed mechanics of that one format. ai-copywriter is channel-agnostic marketing copy: headlines, app descriptions, UI microcopy, email subject lines, and LinkedIn posts, with the humanizing pass built in as a floor under all of it rather than a single dedicated module. Someone writing exclusively for X is better served by x-skills' platform-tuned hook formulas. Someone writing a product's error messages, a landing page headline, and a LinkedIn post in the same afternoon has more use for a single skill that covers all three.

Installing and using it

Installation is a single command: npx skills add mikiarlo3/ai-copywriter --global, or through the Claude Code plugin marketplace. The whole thing lives in one markdown file, SKILL.md, with no code dependencies, so it also works pasted directly into ChatGPT or any other model that takes a system prompt. Once installed, it activates when someone asks Claude to write or punch up copy, drafts against the intake and format rules above, then runs its own draft through the 33-pattern check before returning it.

The honest limitations

The no-fabrication rule is a real constraint, and a useful one: the skill won't invent a benefit or a customer story, so its output is only as sharp as the input it's given. A vague brief still produces vague copy; the intake questions exist specifically to catch that before drafting starts, but a user in a hurry can skip past them. The 33-pattern list also isn't infallible. It's a checklist built from documented tells, not a guarantee that a careful reader won't spot AI involvement, and it can't judge whether a claim is actually true, only whether it sounds fabricated.

The traction picture, verified today

As of today, ai-copywriter sits at 969 GitHub stars and 29 forks, up from 958 stars when it first surfaced on this radar. It was created July 24, 2026, and last pushed August 1. It's MIT licensed, with the humanizer half's copyright held by Siqi Chen and the copywriting additions credited to Mickey Haslavsky, a detail that matters here: the humanizer isn't original to this repo, it's a documented extension of prior work, layered with a copywriting method on top. That lineage, combined with the specificity of the 33-pattern list over a vaguer "write naturally" instruction, is what separates this from the wave of AI-content tools that promise human tone without saying what that means in practice.