skill-creator: Create, revise, and measure a skill

Who it is for

Use skill-creator when you need to create a skill from scratch, revise an existing skill, improve it, run evals, do a performance comparison with variance analysis, or change description to raise trigger accuracy. It is for creating, modifying, improving, and measuring a skill, not for publishing it to an unnamed marketplace.

A better fit is being willing to think through what this skill should do and roughly how, before you enter drafting and testing. The source material does not give an agent-product list or an install command; do not invent a repository path. The official description page is

https://raw.githubusercontent.com/anthropics/skills/main/skills/skill-creator/SKILL.md

First attempt in 3 minutes

The recommended path is not an installer. It is the process in the source material: first judge which step the user is on, then help them move forward. The order is roughly: decide the goal and the approach, write a draft, prepare a few test prompts and run them with claude-with-access-to-the-skill, then help the user read the results from both qualitative and quantitative sides.

Prerequisite: you already have a task you want to turn into a skill. For a low-risk first pass, do not write files. First confirm in a few sentences what this skill should do and roughly how, and judge whether the other person is still choosing a topic, already has a draft, or is already running tests. The summary allows them to say they want to go by feel first and skip evals; if they clearly skip, do not force a test open.

Success signal: you can point to which segment they are in, and the next step advances only one segment. Writing a draft will change files; confirm the storage location before writing. Do not start by running an unnamed evaluation platform.

Usage cases

Case 1: Write a skill from scratch that will trigger. Decide what it does first, then write the draft. description is the main trigger mechanism: write both what it does and in which situations to use it; put every “when to use” into description, not into the body. The source material warns that Claude currently tends to undertrigger, so the description must be specific enough. The done criterion is that description in the draft contains both capability and trigger situations. The excerpt does not give the full field template of SKILL.md, so do not invent required fields that never appeared.

Case 2: After the skill already works, improve triggering. The source material states: after you are done (order can still be flexible), you can run skill description improver, a separate independent script for improving triggers. The done criterion is that you are ready to run this improver, not that every evaluation set is already revised. The excerpt does not write the script filename or command line, so do not invent invocation parameters; only confirm this step: there is another script, and the goal is triggering. Test prompts are run with claude-with-access-to-the-skill, then assessed qualitatively and quantitatively.

Boundaries

The process can be flexible, but the duty is to locate which segment the user is stuck in and advance it. The creation loop is: decide, draft, test prompts, evaluate, revise again. description carries every “when to use”; the body does not restack trigger conditions. Claude currently tends to trigger less often.

Do not invent an installer or an in-repo path. The source material does not give a client list and does not write the full command for description improver. Anything beyond these stated stages, test entries, and description writing is outside this guide. Before writing files or running tests, confirm the path and whether you would overwrite an existing skill.