Ground heuristic audits and design variations in real product evidence
An agent reads your current flow through Figma MCP, pulls comparable shipped flows through Mobbin MCP, scores the flow against named heuristics with linked evidence, and only then proposes variations that explain what they adopt, adapt, or reject.
How to run it
- 01
Connect Claude Code or Codex to Mobbin MCP (claude mcp add mobbin --scope user --transport http https://api.mobbin.com/mcp, or codex mcp add mobbin --url https://api.mobbin.com/mcp) and to Figma MCP, then authorise both in the browser. Mobbin MCP access requires a paid plan.
- 02
Have the agent read the current flow straight from Figma - screenshots plus design context for each screen - so the audit starts from the real file rather than a description of it.
- 03
Ask a scoped research question, not "design a modern onboarding". For example: "Search Mobbin for 12 mobile KYC onboarding flows from established fintech products. Group them by sequence, permission timing, progress pattern, and recovery option. Return the screen links."
- 04
Run the heuristic evaluation: every finding cites a Nielsen heuristic or WCAG criterion, a severity from 0 to 4, the affected Figma screen, and the Mobbin references that show how shipped products handle the same moment.
- 05
Ask for synthesis, not copying: which patterns repeat, where products differ, and why they might. Then add your own context - audience, platform, design system, technical limits, accessibility requirements, and business risk - because references cannot decide for you.
- 06
Generate two or three variations written back into Figma, each with a short rationale naming the patterns it adopts, adapts, or rejects and the findings it resolves, so reviewers compare reasoning rather than aesthetics.
- 07
Keep a human review gate: open the cited sources, test the sequence and its edge cases, and confirm the chosen variation respects product constraints before anything moves to code.
Anti-pattern this prevents
An agent that invents a plausible checkout or paywall from its own assumptions - or a team that spends half a day screenshotting competitors - producing an audit built on opinion and variations nobody can justify.