ai-copywriter: a Markdown skill that writes copy and strips the AI tone out of it
An AI copywriter that uses real copywriting skills + real marketing knowledge with human tone.
At a glance
- What is it?
- mikiarlo3/ai-copywriter packages Wikipedia's signs-of-AI-writing patterns together with an interview-first copywriting method, in a single SKILL.md file you drop into an agent harness. The interesting part is what it refuses to do.
- Who is it for?
- Adopt ai-copywriter if you already run a skill-capable agent harness and you want your headlines and microcopy checked against a fixed list of 33 AI-writing patterns before they ship. Do not adopt it if you need a hosted tool with a UI, or if you cannot supply real product numbers, because the skill asks rather than invents them.
- Can I use it commercially?
- Yes. MIT is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository last received commits 49 days ago.
- What is it written in?
- Mainly Python, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on September 19, 2026, and from our analysis. They are not legal advice.
DEEP OPEN-SOURCE ANALYSIS
The gap ai-copywriter is aimed at: cleanup tools versus writing tools
Most AI writing tools sit at one end of the pipeline. They generate. A smaller group sits at the other end. They take text that already exists and remove the tells: the inflated adjectives, the rule-of-three padding, the sentences that say nothing. mikiarlo3/ai-copywriter tries to occupy both ends in one artifact.
The README is explicit about the lineage. The project is built on blader's Humanizer, which packaged Wikipedia's "Signs of AI writing" guide into 33 detectable, fixable patterns. According to the README, all 33 patterns are still present and unchanged. What this project adds is generation: clickbait titles, short descriptions, microcopy, subject lines, produced in a way that does not trip those same patterns in the first place.
The intended user is someone who already works inside an agent harness and wants copywriting behaviour to be portable across sessions and tools. The runtime artifact is a single Markdown file, SKILL.md. There is no server, no API key, no model of its own. The skill is instructions, and the model you point at it supplies the intelligence.
The interview loop: how the skill forces specificity before it drafts
The mechanism is a two-question frame applied before any writing happens. First: what is the person feeling at the exact moment the line reaches them? The README gives four concrete situations, and they are worth reading closely because they set tone, length and ordering. A headline reaches someone mid-scroll. An error message reaches someone whose task just broke. An empty state reaches a new user worried they are doing it wrong. A subject line reaches someone deleting on reflex. A frustrated reader needs the fix in the first three words; a skeptical reader needs proof before adjectives.
Second: what is the simplest way to explain this? The README states the skill will keep asking what the product actually does until it can be described in the words you would use across a kitchen table. That is a gate, not a suggestion.
To get there, the skill interviews. It asks for three inputs in one batch, skipping anything already supplied: the ICP, the category, and the story. The ICP is defined down to what the person would type into a search box at 11pm. The category is the mental shelf the reader files you on, which determines who you are compared against. The story is the real moment behind the copy, with real numbers.
The part that distinguishes this from a prompt template is the pressure test. Before drafting, the skill checks whether the story contains a surprising number, a moment it almost failed, a belief that turned out wrong, or something you would tell at dinner unprompted. If none of those are present, the README says it keeps digging rather than drafting. Generic inputs get follow-up questions just as missing inputs do. The README lists examples: what do they complain about, in the words they would use, and what claim would nobody else in the category dare to make.
One constraint is stated flatly. The skill refuses to invent product facts. If the strongest headline needs a number, the number has to come from you. That is the design decision that makes the output checkable, and it is also the reason the skill can stall on a user who has no specifics to give.
Installing ai-copywriter with the skills CLI or the Claude Code plugin
The README gives three installation routes. The primary one is the cross-agent skills CLI, run through npx. This installs the skill globally so it is available in every project:
npx skills add mikiarlo3/ai-copywriter --globalAfter running it, start a new agent session or reload skills, because the README notes the skill is not picked up mid-session. To update an existing install:
npx skills update ai-copywriter --globalIf you want the skill registered in every harness the CLI supports rather than the default one, the README gives a wildcard form:
npx skills add mikiarlo3/ai-copywriter --global --agent '*'Dropping the --global flag produces a project-local install instead, which can be committed and shared with collaborators. That is the version to pick if you want the copywriting behaviour versioned alongside the codebase it describes.
Claude Code users have a second route, through the plugin marketplace. Both commands are slash commands typed into the agent, not shell commands:
/plugin marketplace add mikiarlo3/ai-copywriter
/plugin install ai-copywriter@ai-copywriterThe README states the skill is then invoked as /ai-copywriter:ai-copywriter. For any harness that is not covered by the CLI or the plugin system, the manual route is a plain clone into whatever directory your harness reads skills from:
git clone https://github.com/mikiarlo3/ai-copywriter.git /path/to/your/skills/ai-copywriterA first real use, based on the README's description of the interview: ask for a subject line for a product launch and expect to be asked for the ICP, the category and the story before any draft appears. If you answer with a demographic and no moment, expect follow-up questions rather than output. The README also documents a no-install path for chat tools. On claude.ai, create a Project, upload SKILL.md to its knowledge, and put one line in the project instructions telling it to follow SKILL.md for all copywriting and humanizing requests. For ChatGPT, create a custom GPT and paste the full contents of SKILL.md into the Instructions field.
Where ai-copywriter gets in the way
The interview is the feature and the friction. If you want a headline in ten seconds for a low-stakes A/B test, this skill will ask you who the reader is and what just happened to them, and it will keep asking until the answer is specific. That is a poor fit for high-volume, low-consequence copy where speed beats precision.
The refusal to invent facts is the second constraint. A skill that will not fabricate a number is more honest than one that will, but it means the quality of the output is capped by the quality of what you bring. If your product has no surprising number, no near-failure, no reversed belief, the README's own framing suggests the skill will struggle to produce copy that is both true and interesting. That is arguably correct behaviour and still a real limitation for early products with no story yet.
The third issue is verification. The README states that all 33 Humanizer patterns are present and unchanged, but it does not enumerate them in the portion available, and it does not describe how to check that a given output passed all 33. There is no test command, no linter, no exit code. Enforcement depends on the model following instructions, which is probabilistic. Treat the pattern list as a checklist the model is asked to honour, not as a validator that will reject a failing draft.
Finally, the repository's default branch is named claude/humanizer-copywriting-skill-u5x4vd. That is a working branch name, not a release branch, and there are no releases retrieved for this project. If you clone without specifying a branch you get whatever that branch holds at the time.
ai-copywriter against a general-purpose prompt or a hosted copywriting tool
The obvious alternative is a plain prompt: paste your brand voice into a chat window and ask for a headline. The difference is what happens when the model drifts. A prompt has no fixed failure list, so the same model that avoids "unlock your potential" on Monday may produce it on Tuesday. ai-copywriter's approach is to attach a named, finite set of patterns to the request, sourced from Wikipedia's signs-of-AI-writing guide via blader's Humanizer, so the check is at least consistent across sessions.
The second alternative is a hosted copywriting product with a UI, saved brand profiles and a history of past generations. Those tools win on ergonomics and lose on portability. ai-copywriter is a Markdown file. It runs in any harness that accepts skill-style instructions, including chat tools where you paste the file into a project or a custom GPT's instructions field, and it can be committed to a repository and reviewed in a pull request like any other file. There is no account, no dashboard and no usage limit beyond whatever your model provider enforces.
The comparison that matters most is direction of work. Cleanup-only tools assume you already have a draft and want the AI tells removed. Generation-only tools assume you want volume. ai-copywriter assumes you have a real product with a real story and that the bottleneck is specificity, which is why it spends its effort on questions rather than on sentence templates.
Licence, maintenance and the cost of keeping the skill current
The project is MIT licensed, which permits commercial use, modification and redistribution provided the licence and copyright notice are retained. If you fork it to add your own brand rules, the MIT terms travel with the fork. This is a description of the licence text, not legal advice; check the LICENSE file in the repository for the binding terms.
On maintenance: the last push to the repository was on 2026-08-01. There are no releases retrieved for the project, so there is no version number to pin and no changelog to read. Upgrades happen through the skills CLI update command shown above, or by pulling the repository if you installed manually. Because the runtime artifact is a single Markdown file, an upgrade can change the behaviour of every copy request you make, and the only way to see what changed is to read the diff on SKILL.md.
That is the real upgrade cost. There is no semantic versioning boundary between one set of instructions and the next, so a pull can quietly alter how the skill interviews or what it refuses to write. If you depend on stable output, pin a commit and review changes before moving the pin.
Editorial conclusion
Adopt ai-copywriter if you already run a skill-capable agent harness and you want your headlines and microcopy checked against a fixed list of 33 AI-writing patterns before they ship. Do not adopt it if you need a hosted tool with a UI, or if you cannot supply real product numbers, because the skill asks rather than invents them. Verify first that your harness loads SKILL.md the way the README describes, and check whether the branch you clone is the default branch, since the repository's default branch is claude/humanizer-copywriting-skill-u5x4vd rather than main.
Frequently asked questions
Can I do copywriting with AI using ai-copywriter?
Yes. The skill is designed for exactly that: it writes clickbait titles, short descriptions, microcopy and subject lines, and it also removes signs of AI-generated writing from the result. It runs in any harness that supports skill-style instructions, or in a chat tool where you paste SKILL.md into the project or custom GPT instructions.
What is ai-copywriter?
It is a portable agent skill distributed as a single Markdown file, SKILL.md. The README describes it as doing two halves of the copy job: writing copy that earns attention, and stripping out signs of AI writing. It is built on blader's Humanizer and its 33 patterns from Wikipedia's signs-of-AI-writing guide, which the README says are all still present and unchanged.
Which AI is best for copywriting with ai-copywriter?
The README does not rank models. It states that the skill is one Markdown file with no code or dependencies, so any LLM that accepts text can run it, and it documents setup for Claude Code, claude.ai Projects and ChatGPT custom GPTs. The model supplies the intelligence; the skill supplies the method and the pattern list.
Community notes