ai-copywriter's default branch is a generated skill scaffold, and its tree is not one file

The default branch is a scaffold name, not main

The repository metadata names the default branch, and it is not what you would expect from a published project. It is a long branch name that reads like an agent-generated scaffold identifier, combining a tool name, the feature being built and a short random suffix. Anyone running a plain clone with no branch argument gets that tree, not a mainline, and there are no GitHub releases to fall back on, so there is no tag to pin either. The last push to it was 2026-08-01. None of this says the work is unfinished, and the page presents a finished method with a clear method section and complete install instructions. It does mean that if you are evaluating this project rather than using it, the first thing to check is which branch a given download URL actually resolves to.

One Markdown file at runtime, four extra directories in the tree

The page's portability claim is about the artifact, and it is precise about it: the runtime artefact is a single Markdown file, so any harness that accepts skill-style instructions can use it, and the page says the whole skill has no code or dependencies. Three install paths follow from that. A cross-agent skills command installs it globally:

npx skills add mikiarlo3/ai-copywriter --global

a second form targets every supported harness at once, and omitting the global flag gives a project-local install you can commit and share. Claude Code users get a plugin route instead, invoked under its own slash command. And a manual route is a plain clone into a skills directory, which is all the harness needs because the file is the deliverable. The tree is where the claim gets interesting, because alongside the single file sit directories for agents, assets, references and scripts, and the repository's primary language is recorded as Python. So the file you run has no dependencies; the repository around it is not one file.

It questions generic answers, not only missing ones

This is the design decision that makes the skill different from a template, and it is stated in its own paragraph. If what the skill knows about the reader would not surprise a colleague, if it cannot say what is table stakes in the category versus what would raise an eyebrow, or if it cannot write the reader's late-night search query word for word, it comes back with follow-up questions instead of writing around the gap. The follow-ups are concrete: what do they complain about, in the words they would use, and what claim would nobody else in the category dare to make. Then comes the line that matters for anyone evaluating output quality: answers that are present but generic get questioned just as proactively as answers that are missing. Most prompt scaffolds treat an empty field as the failure mode. This one treats a filled-in field with nothing in it as the same failure.

The interview comes before the draft, and the story gets pressure tested

Three things are asked for up front, in one batch, skipping whatever you have already told it: who exactly this is for, down to what they would type into a search box late at night; the category, described as the mental shelf the reader files you on, which decides who you are compared against; and the story, being the real moment behind the copy, with real numbers. Then the story itself gets tested before any drafting starts. The questions are specific: is there a surprising number, a moment it almost failed, a belief that turned out wrong, something you would bring up at dinner unprompted. If not, it keeps digging with you, and the page states the consequence plainly: writing from a weak story produces generic copy that no craft can save. That is an unusually blunt claim to put at the centre of a copywriting tool, and it explains why the interview is not skippable.

Two questions decide the tone, the length and the order

The method reduces to two questions asked before a single word is written. First, what is that person feeling at the exact moment the line reaches them, and the page is careful that this is not a demographic but a moment: a headline arrives mid-scroll and half a second from gone, an error message arrives to someone whose task just broke and who may be blaming themselves, an empty state arrives to someone quietly worried they are doing it wrong, a subject line arrives to someone deleting on reflex. That feeling then decides tone, length and ordering, with two worked rules: a frustrated reader needs the fix in the first three words, and a sceptical reader needs proof before adjectives. Second, what is the simplest way to explain this, judged by whether the product can be described in the words you would use across a kitchen table, and if not, the skill keeps asking what it actually does until it can.

It refuses to invent the number

The argument for doing both jobs in one skill runs through two bad outputs. Ask a model for a headline and you get something like Unlock the Ultimate Guide to Revolutionize Your Workflow. Ask it to tone that down and you get something so flat nobody clicks. The page locates both failures in one cause, the model thinking about the product and its adjectives rather than the reader and their half-second of attention. The positive example is a concrete number, and the negative one is a word that means nothing on its own. Then the part that matters for trust: the skill refuses to invent product facts, and if the strongest headline needs a number, the number has to come from you. It will ask rather than make one up. That is a small rule with a large effect on how much of the output you can publish without checking it.

Thirty-three patterns inherited unchanged from another project

The pattern list is not this project's work and the page says so. It comes from an existing humanizer that packaged a published guide to signs of AI writing into 33 detectable, fixable patterns, and all 33 are carried over unchanged. What this skill adds is the other direction, not cleaning prose after the fact but writing headlines, blurbs and button labels that convert without tripping any of them. The page's claim is stronger than tolerance: the humanizer rules are not a constraint on the copywriting, they are most of what makes it good. For anyone using this, that framing has a practical consequence, because a harness without skill support can still work. Get the file's contents in front of any model that accepts text, tell it to follow them, and the method applies, whether that is a project knowledge base, a custom assistant configuration, or one pasted first message.

Editorial conclusion

ai-copywriter fits someone writing product copy who has been burned by generic output and is willing to answer an interview before getting a draft, particularly if the number in your headline exists and you can supply it. It does not fit you if you want a one-shot generator, because the skill's first act is to ask three questions and then keep asking. Before you install it, check the branch you are cloning, since the default one is not main, and read the additional directories in the tree, because the claim that the whole skill is one Markdown file describes the runtime artifact rather than the repository.

Frequently asked questions

Official sources

1. Issues
2. License: MIT
3. mikiarlo3/ai-copywriter on GitHub
4. README