Use skill-creator to turn the prompt you keep retyping into a permanent skill

A skill is a folder Claude loads when its description matches what you've asked for — a SKILL.md with YAML frontmatter on top, plain Markdown instructions below, and optionally scripts/, references/, or assets/ directories next to it. Skills live in ~/.claude/skills/ globally, or in a project's .claude/skills/ for project-specific ones. The skill-creator skill is the one that builds new ones. You point it at a workflow you've been retyping into prompts, and it interviews you about the intent, writes the SKILL.md, generates test prompts, runs them with and without the skill loaded, compares the outputs, and iterates the description until the skill triggers reliably on the phrasings you actually use.

Why this matters right now. Skills are how the Claude stack becomes composable. Each one is a packaged capability you don't have to re-explain, and the loading model is built so that only the metadata (name plus description, roughly 100 words) stays in context until a skill triggers. Body, scripts, and references load on demand. The implication is that an installed skill you never use costs you almost nothing, so the right move is to install promiscuously and author the ones you can't find. skill-creator is the entry point to that compounding move: once you can author skills cleanly, every recurring task becomes a one-time cost.

Capture the prompt that already worked. The most common entry point. You wrote a careful 600-word prompt last Tuesday that produced exactly the report you wanted; today you can't find the original, so you rewrite it from memory and the output is worse. Ask skill-creator to package it. It captures the intent from your description, writes the SKILL.md, lands the file under ~/.claude/skills/, and the next session triggers it automatically when you describe the task. The prompt becomes a permanent fixture instead of a thing you keep almost-remembering.

Let it tune the description for you. The most failure-prone part of authoring a skill is the description — the line of YAML frontmatter that controls when Claude reaches for the file. Too vague, the skill never triggers; too narrow, only one phrasing works. skill-creator ships with a description-improver script that takes your draft, generates synthetic prompts that should and should not trigger the skill, runs them, and tells you which phrasings need to be added or removed. Faster than guessing, and much faster than finding out three weeks later that the skill is dead weight.

Scaffold scripts and references in the right shape. Skills are not pure prose. A scripts/ directory holds executable code Claude can call when the workflow needs a deterministic step (parse this CSV, hit this API, render this template), and a references/ directory holds long-form docs that load only when the model decides it needs them. skill-creator knows this anatomy and will lay out the right folders based on what you describe — so you don't paste a 4,000-token reference table into the body of SKILL.md when references/api.md would do.

The trap to avoid is over-engineering. Don't start with three reference files and a full eval harness. Start with a SKILL.md that contains exactly the prompt you would have typed by hand, get the description triggering reliably on the phrasings you actually use, and only then add scripts, references, or evals. A small skill that does one thing well and triggers consistently is worth ten skills with elegant directory structures that never fire.

See the try-it block for a one-line invocation that captures whatever you've been retyping this week.