Design-To-UI is a design QA skill for validating UI implementations against Figma designs. It provides probes for detecting layout drift, glyph mismatches, color inconsistencies, edge/border positioning, and text metrics (advance width, weight) between design and implementation.
9 tools with detailed technical descriptions but critical gaps in LLM-friendly structure. All tools have descriptions (50-300 chars, good length), but parameter descriptions are highly technical and assume domain expertise. Input schemas are present with types, but lack enums for constrained values (e.g., 'mode' in color_probe accepts 'dark'|'light' but not declared as enum). No output schemas documented. Tools are read-only (good for safety) but lack error recovery guidance. Naming is verb-based but overly specialized (align_probe, glyph_probe), LLMs may struggle to infer when to use each. No tool annotations present.
Per-block alignment probe — detects block-level position drift by finding the best (dx,dy) shift that minimizes MAE in each grid cell or region, reporting cells where shift exceeds threshold as layout error candidates (padding/offset/gap/letterSpacing issues).
Full-screen capture from Android device/emulator via adb screencap. Ensures device is awake, validates foreground package, detects solid-color (asleep/secure) screens, and waits for framebuffer to settle after wake.
Web capture adapter (single story) — renders specified Storybook story via Playwright/Chromium, captures target element as real.png. Waits for document.fonts.ready for font matching.
Per-text semantic color probe — extracts representative color (darkest or brightest pixel) from each region and compares Euclidean distance to expected hex (Figma token value), flagging mismatches above threshold.
Exact-crop — extracts target window region from full-screen capture. Supports four modes: 'frame' (dumpsys window frame), 'auto' (content-based edge matching), 'anchor' (fixed element anchor), 'box' (direct coordinates). Includes alignment gate to detect crop offset errors.
Edge/full-bleed probe — compares ink pixel start/end coordinates (not area average) for thin lines/borders/dividers. Detects where lines begin/end, immune to anti-aliasing. Flags position mismatches, missing edges, and length discrepancies.
No output schemas documented. LLMs cannot plan downstream calls or extract structured results. Tools return text/JSON but format is not formally specified.
Constrained parameters lack enum declarations. 'mode' in color_probe/glyph_probe accepts 'dark'|'light'; 'emit' in enumerate_regions accepts 'glyph'|'text'|'edge'|'full JSON', all should be enums, not free-form strings.
Parameter descriptions are highly technical and domain-specific (e.g., 'Ink threshold: pixels darker than (bg - ink)', 'Intersection-over-Union (IoU) of normalized ink masks'). LLMs unfamiliar with image processing QA will struggle to use these tools correctly.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 43 | 2026-07-28+ | v2 |
Probe region enumeration — mechanically determines what to inspect from Figma metadata (MCP XML or REST JSON). Classifies nodes as glyph (vectors/small icons), text, or edge (thin full-bleed elements). Outputs deterministic region lists for probes, enabling reproducible coverage.
Glyph identity probe — compares icon shape via Intersection-over-Union (IoU) of normalized ink masks (32×32 grid). Detects wrong glyphs (cross vs arrow). Optional size-check compares original ink bbox dimensions in absolute pixels or dp, immune to anti-aliasing via mid-level threshold.
Per-text glyph width/weight probe — measures ink bbox width ratio (advance/letterSpacing) and coverage ratio (weight/faux-bold) in text regions. Detects systematic letterSpacing errors and faux-bold synthesis via global median deviation analysis.
No error handling or recovery guidance. Tools fail silently or with raw exceptions. No guidance on retryability, user-fixable errors, or next steps (e.g., 'If crop fails, try --mode auto instead of --mode frame').
Tool names are overly specialized ('align_probe', 'glyph_probe', 'edge_probe'). LLMs may not infer that these are quality-assurance inspection tools. Consider grouping under a common prefix (e.g., 'qa_align_probe', 'qa_color_probe') or adding a discovery tool that explains the QA workflow.