Skip to main content

Stunt Double + Claude (design and chat)

Claude is where a growing share of design work starts: interactive artifacts, landing pages, dashboards, and full prototypes sketched in a conversation. The gap is validation. Claude can generate ten variants of a flow, but it cannot tell you which one a distracted first-time user actually completes. With the Stunt Double connector, the same conversation that produced the design can also test it.

Who this is for: designers and PMs designing in Claude (artifacts, Claude design previews), and anyone using Claude as their daily assistant who wants research and QA on tap.

Setup (once)

Pick whichever fits how you use Claude:

  • Claude Desktop, one click: download the extension from stuntdouble.io/integrations (stuntdouble.mcpb) and open it with Claude Desktop.
  • Claude (web or desktop), manual: Settings -> Connectors -> Add custom connector, URL https://app.stuntdouble.io/api/mcp.

The first time Claude uses a Stunt Double tool, a browser window opens to sign in and authorise. No API keys.

Once connected, the workflow prompts (validate_design, run_user_research, triage_feedback, and friends) are available from the prompt picker, and Claude will also reach for the tools unprompted when a testable URL shows up in conversation.

Workflow: design in Claude, validate with personas, iterate

  1. Design as usual. Ask Claude for the artifact: a pricing page, an onboarding flow, a dashboard.

  2. Publish it. Use Claude's publish/share option on the artifact so it has a public URL. Stunt Double actors need a reachable URL; a published artifact is exactly that.

  3. Validate in the same chat. Use the validate_design prompt, or just say:

    "Run this past 3 Stunt Double personas: a rushed mobile user, a sceptical CFO, and someone unfamiliar with the product. Can they work out what we sell and what it costs?"

    Claude creates an interview against the published URL, defines the personas, launches it, and pulls the report back into the conversation.

  4. Iterate with evidence. Findings arrive as themes, task success, and transcript quotes. Because the design lives in the same chat, Claude can apply the changes to the artifact immediately and re-run the identical interview for a before/after comparison.

This loop (generate, publish, test, revise) typically fits inside a single working session, which is the point: validation happens while the design is still cheap to change.

Workflow: research without a design

Stunt Double is also useful in Claude before anything is designed:

  • Concept and comprehension checks against competitors or your current production site: "Interview 4 personas on the competitor's site: where do they get lost in checkout?"
  • Feedback triage: "Triage the new feedback on the marketing site project" pulls in feedback pins (with screenshots and device info), clusters them, and can reproduce suspected bugs with a checklist run.
  • Standing guardrails: "Set up daily checks on signup and checkout" uses the setup_guardrails prompt to create checklists plus a scheduled workflow, so this becomes ambient rather than on-demand.

Tips

  • Personas beat testers. Ask for personas with real texture (background, goals, patience, device) rather than "a user". Variety across participants is what surfaces the interesting failures.
  • Tasks beat questions. "Find the price of the annual plan and start checkout" produces observable evidence; "do you like this page?" produces pleasantries.
  • Keep the interview. Re-running the same interview after changes gives you a longitudinal signal, not a one-off anecdote.
  • Runs take a few minutes; Claude polls and reports when done. Watch the browser session recordings in the dashboard at app.stuntdouble.io when you want to see exactly what a persona did.