respondent-panel

Respondent Panel

Announce at start: "I'm using the humane:respondent-panel skill to collect gut reactions from isolated synthetic respondents."

Gut-level audience reaction to something user-facing, from several people at once, each of whom has seen only the thing itself.

One reaction is an anecdote. Five reactions that all stumble on the same word is a finding. This skill is about getting the second one.

When to invoke

• "How does this tagline land?"
• "What would actual people think of this?"
• "Test this copy / these brand values / this landing page hero."
• "Run a respondent panel."
• After jtbd or brandkit, to check whether the language you landed on survives contact with someone who has not read the reasoning.

When NOT to invoke

• The artifact is a document meant to be studied — a PRD, a spec, a pitch deck script. That is persona-review: expert stakeholders reading carefully and arguing back. This skill is strangers glancing.
• You want a usability judgement of an interface. That is nielsen-heuristics.
• You want the copy fixed. Respondents deliberately do not rewrite; bring their reactions back and revise yourself — ux-writing owns the rewrite, and knows to revise against convergent findings only.

The one rule

Respondents see the artifact and nothing else.

Not the brief, not the JTBD corpus, not the positioning rationale, not the six drafts you rejected, not what you were hoping they would feel. Every sentence of context you add buys you a more agreeable answer and a less true one.

This is why respondents run in isolated contexts — a respondent that shares your session has already read everything you know and cannot un-read it.

Claude Code extras: launch each respondent as the bundled synthetic-respondent agent, all in a single message so they run concurrently and independently. Pass the artifact verbatim plus that respondent's persona brief — nothing else.

On other agents: run them sequentially in fresh sessions (or after clearing context), pasting only the persona brief and the artifact. If neither is possible, run one respondent and say plainly in the output that it is a single uncontaminated reaction, not a panel.