Working in Claude Code

The recommended way to generate skills is from inside Claude Code. Claude reads each page's code in your session, runs skillgendev's commands, follows each result's next step and asks you only what needs a decision.

Install the integration

skillgendev install                       # into ~/.claude
skillgendev install --target ./.claude    # into this project only
skillgendev install --no-chromium         # skip the Chromium download

install copies three files from the package's claude/ folder, with the tool name filled in:

File	What it is
commands/skillgendev.md	The /skillgendev slash command
agents/skillgendev-route-analyzer.md	A subagent that analyses one route with read-only tools (Read, Grep, Glob, Write)
skills/skillgendev/SKILL.md	The driver skill: how to run skillgendev and follow its next steps

When the probe extra is installed, install also downloads Chromium for --xy. Rerun it after updating skillgendev; it copies the files rather than linking them.

Any agent can drive skillgendev

The driver instructions are plain Markdown (claude/skills/driver/SKILL.md in the repository), so any agent that can run commands can follow them. install only sets them up for Claude Code.

Use /skillgendev

/skillgendev /campaigns/new
/skillgendev all
/skillgendev all --name acme

What the command does, in order:

1. If .skillgendev/config.json does not exist, it runs skillgendev init and asks you for the app's display name and, if you have one, a dev URL.
2. For all, it runs skillgendev routes --all --json, shows you the routes, and lets you drop or rename any before generating.
3. It runs skillgendev generate --route <route> --json (or --all), passing --name through, and --xy when you want field positions.
4. It follows each result's next until nothing is left:
   • analyse: Claude reads the prompt and writes the analysis itself, using one route-analyzer subagent per route when there are several.
   • ask: Claude asks you the question in chat, runs run on a yes, then runs then.
   • status exists: Claude asks before a skill is recreated.
5. It reports each skill folder and any open questions, asks you those questions, records your answers with skillgendev answer, and generates again with --overwrite.

How the analysis runs

There are three analyzers, set with --analyzer or in config.json:

Analyzer	Who reads the code	Used by
claude-code (default)	skillgendev runs headless claude -p with read-only tools (Read, Grep, Glob) and a JSON schema, one route per run	generate in a terminal or CI
file	Claude Code writes analysis.json itself in your session	interactive /skillgendev
replay	recorded analyses named <route-slug>.json	tests and fixtures

If skillgendev cannot call Claude Code itself (not installed, not signed in), the run does not fail. It writes the prompt, and the result carries next.action: "analyse" with the prompt and the path to write. The agent running skillgendev does the analysis and runs the command again. See The agent handoff.

If Claude's own usage limit is reached, the result says so with need: "limit". The agent tells you and stops, rather than reading the code by hand, which would spend far more of the same limit. Run then once the limit has reset; finished analyses are reused.

prepare: just the prompt

skillgendev prepare --route /campaigns/new

prepare builds the page slice and writes the analysis prompt to .skillgendev/pages/<skill>/prompt.md without generating anything. /skillgendev runs it for you; you rarely need it by hand.