FAQ

What does skillgen produce?

One Agent Skills folder per page of your app, with a SKILL.md that tells an AI what every field means, what it accepts, the fill order, the rules, and how to fill and check the page. See Anatomy of a skill.

Which AIs can use the skills?

Any AI that accepts Agent Skills and has a browser: Claude, Cowork, ChatGPT and others. Each skill carries four fill methods, so it works whether or not the AI's browser can run JavaScript in the page.

Does it need an MCP server, an extension or an API key at use time?

No. Skills use the AI's built-in browser and nothing installed. The library never depends on or mentions MCP servers.

Which frameworks are supported?

Any web app whose forms render as HTML in a browser. Frameworks matter only for finding routes; see Route discovery. When no hint pack matches, Claude Code reads the code graph to find routes.

Do I need a running copy of my app?

No. Generation reads code only by default. A dev or staging copy is needed only for --xy, which adds field positions and a layout check.

Is it safe to point the probe at my app?

Only at dev or staging. The probe refuses an environment marked production, intercepts and cancels write requests, never submits, and never clicks things like Save, Delete or Sign out.

What does generation cost?

Only the analysis step spends model tokens: Claude Code reads up to about 60k tokens of code per page. The spec estimates roughly 0.1 to 0.3 million tokens processed per route. Everything else is deterministic and runs on your CPU. Analyses are reused while a page's code is unchanged.

How much does a skill save at fill time?

The design spec models a 40-field form at about 8 tool calls with a skill against about 138 with computer use, roughly 25 times fewer input tokens. That is a model from the spec, not a published measurement yet.

Can I rename a skill?

Yes. You always choose names: --name on generate, or edit skill in .skillgen/routes.json. Names must be lowercase letters, digits and single hyphens, at most 64 characters.

Will regenerating erase my edits?

Not the ones between <!-- skillgen:user:start --> and <!-- skillgen:user:end -->. Everything else in SKILL.md is generated and is rewritten. Put your notes and rules in the user section, or use skillgen add-rule.

Is there a verify command?

Not yet. The spec's with-and-without-skill runs would need a browser in headless Claude Code, which would mean an MCP server, so it is an open question in the decision log. Each skill ships evals/evals.json, and the repository's test bench has 14 manual cases.

Is it on PyPI?

Not yet. Install from a clone; see Installation.

Why is it called skillgen?

It is a working name. The name lives in one constant (skillgen.TOOL_NAME) so it can be changed in one commit when a final name is chosen.

What licence is it under?

Apache License 2.0.