Field positions with --xy

generate has two modes, and the command chooses:

Command	What you get
skillgendev generate --route /x --name my-skill	Normal. Everything from the code. No browser, no x/y. Works on every page
skillgendev generate --route /x --name my-skill --xy	Everything normal gives, plus x/y. Opens the page on your dev copy in Chromium and measures where each field sits

--xy only adds. It never removes a field, changes what the skill says about one, or adds a question. If it cannot measure the page, you still get the normal skill, and the AI running skillgendev is told why.

Requirements

• The probe extra (uv tool install "skillgendev[probe]") and Chromium (skillgendev doctor --fix).
• A dev or staging copy. Never production: the probe refuses an environment marked production, intercepts and cancels write requests, and never submits anything.

Set the address once, or pass it per run:

skillgendev init --url http://localhost:3000
skillgendev generate --route /campaigns/new --name new-campaign --xy
skillgendev generate --route /campaigns/new --name new-campaign --url http://localhost:3000   # implies --xy

If you forget the address, the run writes the skill from the code and the agent asks you for it, then runs the command again.

What the run measures

The skill is always built from the code first. --xy then opens the code-built page, makes any open steps, and measures every field by the skill's own selector. That means a measured box always belongs to the element the skill will fill, and the measuring function can only return a layout, never change a field.

The run reports how many fields got x/y (for example x/y for 16 fields). When x/y cannot be measured, x/y from an earlier run is kept for every field whose selector has not changed. It only ever verifies, so an outdated box can never aim a click.

What ships with the skill

Every measured position ships in references/layout.json: the window size, each field's x/y, width and height, and its offset from the first field, plus which view it was measured in. The AI reads it only when it has to check the page by hand, and measures a field on screen again before clicking it. export copies it with the skill.

The layout check at fill time

With a layout, the fill call first checks the live page against up to eight landmark fields (one per row), compared by structure rather than pixels, so a different screen size is fine:

• match or resized: the form is filled.
• blocked (a dialog covers the form) or changed (not the layout expected): nothing is written; the AI takes a screenshot and inspects the page as usual.

The check is a shortcut, never a gate, and adds roughly 1,250 tokens to each fill.

Signing in

If the page needs a sign-in, you do not have to know that beforehand. The skill is written from the code anyway, and the agent asks whether to open a browser window so you can sign in:

skillgendev login --url http://localhost:3000/login

You sign in and close the window. The session is saved in .skillgendev/secrets/storage-state.json (gitignored, never exported), and the next --xy run uses it.

Looking at one page: probe

skillgendev probe --route /campaigns/new
skillgendev probe --route /campaigns/:id/edit --url http://localhost:3000/campaigns/7/edit

probe opens one route on the dev copy and shows what the browser finds there: which fields, which values it can set, what a submit would send (keys and types, never values). It saves the capture to .skillgendev/pages/<skill>/probe/ and does not change a skill. Use it to see what the probe sees.

What can stop --xy

Need	What the agent does
url	Writes the skill from the code, then asks you for the address
server	Writes the skill from the code, then asks you to start your dev server
chromium	Writes the skill from the code, then asks whether to download Chromium (about 150 MB)
probe	The probe extra is not installed; asks you to reinstall with it
login	Writes the skill from the code, then asks whether to open a sign-in window
open	No field showed. First tries the clicks the code points to; asks which button opens the form only if none works

Each of these is a next step. Nothing is analysed twice while the page's code is unchanged.