Free Claude skill
Turn a task you keep doing by hand into askill file Claude follows on its own, and test it before you trust it.
It writes a complete SKILL.md, tests it on three examples and fixes what fails. The tests are a check, not proof.
Free with an account. No password: you get a login link by email, and it stays in your library.
If you have not named one, it asks what you keep asking Claude to do, the same way, again and again. If you name several, it asks which to build first. One skill per task.
Two messages at most. First: what you give Claude each time, what you want back (format, length, sections, tone, language) and one or two real past examples, with names, passwords, card or ID numbers and customer data removed. Then only what is still unclear: rules, what it must never do, the words you type to trigger it, what it is not for and where you use it. If you give no triggers it suggests three, marked suggested. It never guesses about your work, data, company or audience: it writes [ADD: ...] and lists each one at the end.
One complete file of at most 150 lines in the official format: a name that matches the folder, a description with when to use it and what it is not for, numbered steps, an exact output format and rules. English unless you ask otherwise. Your examples shape the output format, but your private data is not pasted into the file. You see the full file and a three line summary.
Before running, a checklist of 5 to 8 yes or no checks taken from your answers. In Claude Code a subagent gets only the SKILL.md and the test input, never this conversation. Without subagents Claude plays the test itself and says the tests are simulated. You see only the result lines.
For every FAIL it gives the cause in one line, changes the SKILL.md (only the changed lines are shown) and runs all three tests again, with before and after side by side. If something still fails after two fix rounds, it stops, explains and asks you before trying more.
It gives the folder layout and the install steps for where you use Claude, or both if unsure. In Claude Code it offers to write the file to disk and does so only when you say yes.
A complete SKILL.md and three test results. This is the shape the results and the file come in.
The test lines
The SKILL.md it writes
Passing three examples does not prove the skill is right, so try it on your next real case. It never invents facts about your task, your numbers, customers or results: it asks, or leaves a placeholder. It never puts passwords, API keys, card or ID numbers or customer data in a skill file, and it tells you to redact them.
[ADD: who the report is for]
One task per skill: if the scope grows, it suggests a second skill instead. It is not for editing Claude Code settings or hooks, building MCP servers or plugins, or writing one off prompts.
skill-smith folder in ~/.claude/skills/ for every project, or in .claude/skills/ inside one project./skill-smith, and name the task you keep repeating.unzip ~/Downloads/skill-smith.zip -d ~/.claude/skills/Best here. A subagent tests the file and sees only the SKILL.md and the test input.
There are no subagents here, so Claude plays the tests itself and says they are simulated.
The skill you build is a file you own. Update it when the task changes.