skillgen

skillgen is a framework- and domain-agnostic compiler. It reads an app's code route by route and writes one build-ready Agent Skills folder per page, so Claude, Cowork or any AI that accepts Agent Skills can fill that page's form with few tokens.

Status: pre-alpha

skillgen is a working name and the project is pre-alpha. It is not on PyPI yet, so you install it from a clone. Interfaces may change between commits; the decision log records why.

Why it exists

Without a skill, an AI rediscovers a form on every task. It takes screenshots, clicks field by field, and guesses which options exist and which fields depend on which. A generated skill tells the AI once what every field means, what it accepts, the order to fill it in and how to check the result. The AI then fills the whole page in one browser call.

The work splits into two phases with different people:

Phase	Who	What happens
Generate	The developer, once per code change	skillgen finds the routes, picks out the code that builds each page, has Claude Code read it, and writes one skill folder per page. A running dev copy is optional.
Use	Anyone the skills are shared with	They add a skill to Claude, Cowork, ChatGPT or another AI. With its built-in browser the AI opens the page and fills it. Nothing is installed and no MCP server is involved.

skills/acme-campaign-new/
  SKILL.md                    # fields, rules, dependencies, and how to fill them
  scripts/fill.js             # the fallback engine, read only for fields the skill cannot do
  references/fill.md          # the one JavaScript call that fills the page
  references/locators.md      # each field's role, label, selector and action
  references/manual-steps.md  # the form as plain steps by visible labels
  evals/evals.json            # four synthetic test requests

Where to go next

InstallPython 3.10+, uv and git. One command puts skillgen on your PATH.Quick startFrom an app repository to your first skill folder in a few commands.Use a skillHow an AI picks a skill and fills a page, and the four fill methods.How it worksRoutes, slices, analysis, merge, generate, validate, and the optional probe.Command referenceEvery command and flag, with its --json output and exit codes.ContributeGround rules, setup, tests, and how to improve these docs.

Principles

These rules come from the project's design and are enforced by its tests:

• Framework-agnostic. Frameworks matter only for finding routes. Router knowledge lives in YAML hint packs; deleting them leaves a working tool.
• Domain-agnostic. No business vocabulary is built in. Field meaning comes from the app's own code, labels and validation.
• One route, one skill. Each skill is self-contained and can be shared alone.
• Code first. A skill is always built from the code. The Playwright probe is an optional extra that only adds field positions (--xy).
• Facts carry evidence. Every fact records its source and confidence. Uncertain facts become questions, and evidence never reaches a skill.
• Agents run the commands. No command prompts or waits for input. A run that needs you says what it needs, and the agent asks you in chat.
• No MCP anywhere. Skills work with the AI's built-in browser and nothing installed.