Let Claude Write Your Skills
You don't need to learn markdown formatting, YAML front matter, or the folder structure a Claude skill expects. You need to describe the job to Claude the way you'd explain it to a new hire — and let a skill that builds skills do the rest.
The problem with hand-writing SKILL.md
A Claude skill is just a SKILL.md file (plus optional supporting files) with a name, a description, and instructions. Simple in theory. In practice, most people who try to hand-write one get stuck on the same things: what goes in the description versus the body, how specific the trigger wording needs to be, whether to bundle examples inline or as separate files, and how to know if the thing actually works before shipping it.
None of that is exotic knowledge — it's just knowledge you have to accumulate by trial and error, one broken skill at a time. That's the gap skill-creator closes.
What skill-creator actually is
Skill-creator is an official Anthropic skill: a skill whose job is to build other skills. Instead of opening a blank file and guessing the structure, you tell Claude what you want in plain English — "I want a skill that edits my newsletter drafts" — and skill-creator takes it from there.
It works as an interview. Claude asks you one question at a time: what should trigger this skill, what does good output look like, are there edge cases it should handle differently. You're not filling out a form or memorizing a schema — you're having a conversation about a job, the same way you'd onboard a new team member.
Once it has enough to work with, skill-creator drafts the skill, tests it against the kind of input you described, and saves it into your skills directory in the correct structure. You never touch the raw file.
How to run it
- Install skill-creator once. It ships as one of Anthropic's official skills, alongside the rest of your Claude Code / Claude Desktop skill set.
- Describe the job, not the format. Say something like: "Use the skill creator to build me a skill that edits my newsletter drafts before I publish them." You don't specify sections, YAML keys, or file layout — that's the tool's job, not yours.
- Answer the interview. Claude will ask targeted questions: what should trigger the skill, what should it never do, what does a good result look like. Answer like you're briefing a colleague.
- Paste real examples — plural. This is the step people skip, and it's the one that determines whether the skill generalizes or breaks the second time you use it. One example teaches the skill to overfit; several examples teach it the actual pattern.
- Let it draft, test, and save. Skill-creator writes the SKILL.md, runs a test pass, and saves the finished skill — ready to trigger the next time the situation matches.
Why the "plural examples" trick matters
The instinct is to describe what you want abstractly — "professional but warm tone," "cut the filler." Descriptions like that are ambiguous even to a human editor. Concrete before/after examples remove the ambiguity: skill-creator can see the actual transformation you're asking for, not just your adjective for it. Two or three examples pinned to different situations (a short update, a long-form piece, a correction) also protects against the skill locking onto one narrow pattern and failing outside it.
Where this pays off
The same describe-interview-draft-test loop works for far more than editing. A few common uses:
- A new writing style — brief it once with real samples, reuse it on every future draft.
- A research audit — describe what "good" research output looks like, and let it check future work against that bar.
- A recurring report — the format and checks stay consistent even as the underlying data changes each time.
In every case, the pattern is the same: you're the domain expert who knows what good output looks like, and skill-creator is the mechanism that turns that knowledge into a reusable skill without you ever opening a markdown spec.
Power tricks
- Feed it real examples, not vague descriptions. Two or three concrete before/after pairs beat any amount of adjective-heavy prose.
- Cover different situations in your examples. A short update and a long-form piece teach a different lesson than two similar samples do.
- Let the interview drive. Resist the urge to hand-edit the resulting SKILL.md — go back through the interview or ask Claude to revise it instead, so the file stays in the structure the tool expects.
Resources
Building an AI feature? Yeda AI designs, audits, and ships production LLM systems.