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.

Workflow

Step 1 — Get the artifact

Ask for the exact text or file to react to. Take it verbatim. Do not tidy typos, expand abbreviations, or add the surrounding context you think is missing — if it is missing for respondents, it will be missing for real readers, and that is itself the finding.

Note the medium, because it sets the encounter: a billboard glimpsed at speed, an app-store subtitle, a landing-page hero, a cold email subject line. Tell each respondent where they are seeing it.

Step 2 — Build the panel

Default to 5 respondents. Three is thin; beyond seven you are paying for repetition.

Vary them along axes that plausibly change the reaction to this artifact. Pick 3–4 axes and make each respondent specific:

Axis	Why it moves the reaction
Relationship to the category	Newcomer vs. burned-before vs. current happy user of a competitor
Ad exposure	Someone who sees forty pitches a day is bored where a rare viewer is curious
Age / life stage	Changes what references land and what reads as dated
Place & language background	Idioms and wordplay travel badly; non-native readers catch ambiguity
Buying power over this	Someone who would pay reads the claims differently than someone who wouldn't

Write each brief as a person, not a segment — two or three concrete sentences. "38, runs a two-van plumbing business outside Leeds, has bought three scheduling apps and abandoned all of them, reads nothing about software" beats "SMB owner, skeptical."

Ask the user to confirm the panel before spending on it, and say which axes you varied and why. If the artifact is aimed at a specific audience the user has already described (a jtbd.json persona, a stated target market), build the panel around that audience rather than a generic public — but keep at least one respondent from outside it, because that is who tells you when the copy only works for insiders.

Step 3 — Run them

All respondents get the identical artifact and medium. Only the persona brief differs. Never tell a respondent what the others said.

Step 4 — Read the panel

The panel is not a vote. Do not average the reactions or declare a winner. Report:

1. Convergence — anything two or more respondents independently hit. Same word misread, same confusion about what is being sold, same comparison to another brand. This is the strongest signal the method produces; lead with it and quote the respondents directly.
2. Divergence — where they split, and along which axis. If the newcomers liked it and everyone who has used a competitor was suspicious, that is a much more useful sentence than "reactions were mixed."
3. Comprehension — how many understood what is being sold, in one pass, without help. Report this as a count. It is often the real finding and it is easy to lose under the more colourful reactions.
4. Dead spots — parts of the artifact that no respondent mentioned at all. Copy nobody reacted to is copy nobody read.
5. What actually landed — the specific phrases that worked, named. A panel that only reports problems will get you a rewrite that loses the good parts.

Quote respondents verbatim rather than paraphrasing them into marketing language. The unpolished phrasing is the evidence — "I thought it was an insurance thing" survives translation into "brand-category confusion" badly.

Step 5 — Hand off

State clearly what this panel is and is not: synthetic reactions that surface confusion, clichés, and tone problems cheaply and early. They are a rehearsal for contact with real people, not a replacement for it. Never present panel output as market research, and never attach a confidence percentage to it.

Then offer the next step:

• Revise the copy with ux-writing against the convergent findings, then re-run the same panel — same briefs, so the comparison is clean.
• Run before-after if the artifact is claiming a transformation.
• Take the convergent findings to real users, if any are reachable.

Output

Markdown. Convergence first, then divergence by axis, then comprehension count, then dead spots, then what landed. Full individual reactions go at the end, under a heading, so the reader meets the pattern before the anecdotes.