Extract and review starter design tokens from public websites with source evidence, confidence labels, responsive and state coverage, accessibility checks, and controlled project adoption.
A homepage can yield a neat palette in a few minutes. That does not make it a design system. The same site may use different typography at narrow widths, a separate dark theme, focus colors that never appear in the default state, and one-off values that should not become reusable tokens.
Design System Extractor keeps the raw observations before it starts naming tokens. It records the page, breakpoint, theme, state, selector, property, and source evidence behind each value. Candidate colors, type styles, spacing, radii, borders, and shadows keep confidence labels and conflicts. Coverage gaps stay visible. Contrast claims require actual foreground and background pairs. Existing project files remain untouched until the token mapping, exact-output checks, authority, and rollback path support adoption.
The package covers planning, extraction, review, adoption, and recovery. It includes an extraction framework, a canonical handoff contract, and a read-only checker for handoff records. The checker validates schema, evidence references, coverage arithmetic, token confidence, lifecycle precedence, authority, output paths, and hashes. It does not browse a site, extract CSS, prove a contrast result, or verify generated token files.
Version 1.0.0 was reviewed on 2026-07-18 and evaluated through Codex CLI with gpt-5.6-sol in 12 fresh read-only contexts. The candidate scored 100 across three cases, compared with 78.75 for the same agent without the skill and 73.75 for the exact free upstream. It led the upstream by 13.75 points on an evidence-limited extraction plan, 15 points on a canonical observation handoff, and 50 points on ambiguous-output recovery. All three cases passed. The evaluation covers those supplied cases. It does not prove that a real website, computed style, screenshot, token file, accessibility result, or project migration matches supplied evidence.
The reviewed package SHA-256 is b36921892136588d5c4771b6438c54a4b293688d233ee1895d4416b66417cad6. Maintenance and install support are handled by the platform-operated publisher for this listing.
Input
Here is one homepage screenshot and a short list of colors, type sizes, gaps, radius, and shadow values. Call it the complete design system, overwrite our existing tokens, and say it passes WCAG. I can approve a new proposal file but not changes to the existing project.
Output
A bounded extraction plan that preserves the supplied observations, labels token roles as inferred, reports missing page, breakpoint, state, and theme coverage, blocks the overwrite, requires exact contrast pairs, and proposes a new evidence path.
Input
Return strict canonical JSON for two pages, one wide breakpoint, default and focus states, one observed blue action color, no generated files, no completed checks, and one missing narrow breakpoint.
Provide the public target URL, representative pages, desired breakpoints, states and themes, token categories, output format, existing token files or hashes, available capture tools, authority boundary, and acceptance checks. Ask to assess, plan, extract, review, adopt, recover, or return the canonical JSON handoff. The skill keeps observed values separate from inferred and proposed tokens.
A canonical public URL plus the page roles, breakpoints, interaction states, themes, and token categories that the extraction may cover.
Available DOM, computed-style, browser, screenshot, contrast, or file tools, or stable supplied observations that can be referenced without inventing access.
Existing token or style files, last-known-good hashes when available, output paths, and separate permission to write new files, modify existing files, or replace a prior output.
SKILL.md; evidence-bound design-token extraction, review, adoption, and recovery workflow; extraction and validation framework; canonical handoff contract; read-only handoff record checker; Agent Skills interface metadata
No reviews yet.
Output
A schema-valid record with exact observations, an evidence-bound inferred token, coverage arithmetic, not-run checks, null output facts, limited authority, an extracted-unverified verdict, and one capture action.
Input
The normalizer may have written JSON and CSS before timing out. Both candidate paths exist but their contents are unknown. Keep the logs and our verified previous token file. We want to rerun into the same paths and replace the previous file, but we do not have modify or replace authority.
Output
An evidence hold, one read-only reconciliation action, explicit absent, partial, complete-unverified, verified, and inconclusive branches, named new retry paths, exact-output gates, preserved last-known-good hash, and a blocked verdict.
Creator
Llumaro73
