The page model

Everything known about one route's form is merged into one internal document, .skillgendev/pages/<skill>/page.json (schema src/skillgen/schema/page.schema.json). Skills are generated from it, and it never leaves your machine.

What it holds

Part	Meaning
route, skill, title, purpose	The page, its skill name and what it is for
app, source_commit, version	The app, the commit it was made from, and the model version
sections	Visible groups of fields (a fieldset, a card, a tab)
fields	Every editable control, keyed by a stable field key
fill_order, dependencies	The order to fill in, and which field controls which
groups	Repeating rows, such as several ad creatives or audience rules
rules	Rules a value must follow, in words, optionally as a machine check
gotchas	Advice the AI should know, each with its provenance
submit	The submit control, the success signal and where field errors show
params	Route parameters such as id in /campaigns/:id/edit
open	Clicks that show a form hidden behind a button
views	Parts shown only after clicks: wizard steps, tabs, dialogs
layout	Measured field boxes, with --xy only
verified	true only when every required field was proven on the running page

Each field carries its label, what it means, its type, whether it is required, any required_if or visible_if condition, its options or where options load from, limits (length, range, pattern, file types), the selector used to find it, and its fill strategy.

Fill strategies

Strategy	Used for
native-input	Text, number, email, date and similar inputs, textareas
native-select	<select>, single or multiple
choice	Checkbox and radio groups
aria-combobox	Searchable dropdowns and autocompletes: type, wait, pick by visible text
aria-tree	Tree pickers: expand nodes along a path, pick the item
aria-listbox	Custom lists and chip pickers
rich-text	Content-editable editors
file	Uploads; needs the browser tool's file-upload action
steps	Anything else, handled by the user or the browser tool

file and steps fields cannot be completed by the fill script on its own; the skill tells the AI how to handle them.

Value types

text, long-text, number, date, email, url, phone, enum, multi-enum, boolean, file, path (a tree path). Conditions compare a controlling field with the operators in, not_in, checked and filled.

Sources

Every fact records where it came from:

Source	Meaning	Base confidence
RUNTIME_PROBED	Seen on the running page by the probe	0.90
EXTRACTED	Stated literally in code (an attribute, a validator entry)	0.85
CLAUDE_INFERRED	Claude's reading of the code, not stated literally	0.60
DEVELOPER_CONFIRMED	Answered by the developer	1.00
USER_DEFINED	A rule added by hand or with add-rule	1.00

Confidence

The arithmetic lives in src/skillgen/model/confidence.py:

• Agreeing sources combine as 1 − (1 − a)(1 − b). Code plus probe gives 1 − 0.15 × 0.10 = 0.985.
• Disagreeing sources cap the fact at 0.5 and raise a question.
• Facts at 0.7 or above go into the skill. Below that they wait for the developer.

So a fact Claude only deduced (0.60) never reaches a skill on its own; it becomes a question, unless the probe or the code confirms it. Descriptive facts (what a field means, its format, notes) never raise questions.

Your answer turns a fact into DEVELOPER_CONFIRMED at 1.0. See Questions and rules.

Evidence

Beside each fact, skillgendev keeps evidence: the subject (such as field:email.required), the source, a location (a file and line, or a probe capture id) and short detail, in .skillgendev/pages/<skill>/evidence.jsonl. Evidence never reaches SKILL.md, fill.js or anything export writes.