Our directory has evaluated 745 skills that other people wrote. Starting today, you can create your own — and it gets the same scorecard before you ever download it.
The most common feedback we get is some version of "I love the directory, but I want a skill for my own workflow, and nobody's written it." A skill for your particular ledger format. Your team's release checklist. Your reading routine.
So we built Create a Skill — a guided wizard that turns four plain-language answers into a working, evaluated SKILL.md. No code, no markdown knowledge, no GitHub repo.
How it works
Step 1 — What should it do? One or two sentences in your own words. "Log my daily expenses and show the running total." Then pick a type: query (looks things up), generate (produces an artifact), or operate (performs actions).
Step 2 — How will you call it? The exact phrases you'd actually say to trigger the skill. This is the single most important input — our data shows trigger quality decides whether a skill ever fires. The wizard requires at least two phrasings and bakes them into the skill's description verbatim.
Step 3 — Does it touch anything? Read-only, changes things, or money-and-production. The last two automatically add mandatory confirmation rules — the skill must list what it's about to do and wait for your approval before acting.
Step 4 — What do you want to see? Optional. Skip it for a sensible default.
Then the AI writes the SKILL.md: frontmatter description with trigger phrases and boundaries, workflow steps an agent can execute directly, explicit boundaries, output format, and — for action-taking skills — confirmation and failure handling. The whole thing stays under 120 lines, following the same structure we see in our highest-scoring skills.
The part that matters: it gets scored
Here's what makes this different from any "AI generates a file" tool: the output is immediately evaluated on the same six-dimension rubric we use for all 745 directory skills — trigger quality, structure, workflow, content, engineering, and security, with the security veto that caps confirmed-unsafe skills at 39/100.
You see the score before you download. If the workflow section scored 12/25, the suggestions tell you why — and you can regenerate with sharper answers. It's a feedback loop that teaches you what good skill-writing looks like, using your own skill as the material.
Take it home
The result is yours:
- Download SKILL.md or a ready-to-install .zip
- Drop it in
~/.claude/skills/<name>/and restart your session - Optionally submit it for review — our editorial queue may list it in the directory (no guarantees, and we check for unsafe patterns before anything goes public)
Your last five creations stay on the page for re-downloading, re-evaluating, or iterating.
Why we built it this way
Three design choices worth explaining:
The wizard asks about triggers, not features. Most people describe what a skill does; almost nobody describes when it should fire. Since trigger quality is the weakest dimension across our whole directory (7.74 average), we made it the one mandatory, structured input.
Risk level is a first-class question. The difference between a skill that reads your ledger and one that writes to it is the difference between "run freely" and "confirm every step." Making you choose upfront means the confirmation rules are written into the skill, not bolted on later.
Evaluation is free and unlimited-ish. Scoring your own creation against the same rubric as the directory's best skills is the fastest way to develop taste for this format. We cap generations per hour to control costs, but evaluating what you've generated costs you nothing.
The feature carries a TEST badge while we tune the generator. Try it here — and use the feedback link on the page to tell us what breaks. The best bug reports will shape what this becomes.
Top comments (0)