Measuring AI implementation quality of Figma designs through controlled experiments.

Setup — Install and configure canicode (CLI, MCP token, Claude Code skills) Decision Log — Every major decision with context and rationale Rule Reference — Complete detection conditions, subTypes, and exemptions for all 16 rules Scoring Model — How scores are calculated, severity levels, grade thresholds Calibration — Pipeline structure, agents, debate highlights, fixture analysis Delta-driven Pipeline — How calibration + rule discovery close the feedback loop Calibration 01 — First full calibration: 10/24 fixtures, responsive comparison, avg 93.7% similarity Calibration 02 — size-constraints responsive strip: 6/24 fixtures, missing-size-constraint -5→-8, raw-value -3→-4

• design-tree @500px: 95% similarity (root width fix only, flex expands naturally)
• MCP Vite @500px: 92% (good, hardcoded widths), 75% (bad-structure)
• MCP requires LLM post-processing for responsive — hardcoded w-[375px] in child elements
• design-tree is inherently responsive — root width removal is the only change needed
• Structure (auto-layout) is the only category that affects responsive — token, component, naming are viewport-independent

Help teams implement designer-provided Figma designs exactly as designed, with zero unnecessary AI token cost.

See #121 for full direction statement.

OAT (One-at-a-Time) Sensitivity Analysis — remove one design property at a time and measure the impact on AI implementation quality.

• #108 — Ablation calibration signals
• #115 — Fixture design spec
• #117 — Input method comparison
• #118 — design-tree only for implement
• #119 — Category weight adjustment
• #120 — MCP get_metadata limitation
• #121 — Project direction update