Persona review

/persona-review

Goal

Answer one question with evidence from a diverse panel of ordinary users: how would real people react to this? You seat five synthetic-but-grounded Korean personas — sampled from Korean census distributions — in front of a target artifact and report their reactions, frictions, and what to change.

This is a citizen/user lens: comprehension, trust, tone, and clarity. It stands beside code-reviewer (correctness) and security-reviewer (vulnerabilities) and replaces neither. A panelist may voice a worry that is really a security or correctness concern ("is my data safe here?") — capture it as a user worry and route it to the right reviewer; do not adjudicate it here.

Leading word: panel. Five independent voices, then one synthesis — not one voice pretending to be five.

Inputs

• Target — the artifact under review: a URL, a file path, pasted copy, a screenshot description, or a component. Required.
• Question (optional) — what to probe: first-read comprehension, trust, tone, call-to-action clarity, pricing legibility. Default when unstated: first-time-visitor comprehension + trust.
• Catalog — skills/persona-review/personas/catalog.json (shipped). If it is missing or unparseable, stop and say so — do not invent personas.

Steps

1. Load the catalog

Read skills/persona-review/personas/catalog.json and confirm personas is a non-empty list. Each entry carries age, sex, province, occupation, education_level, a one-line summary, hobbies / skills, and a review_lens tag (1-2 of UX / 카피 / 접근성 / 신뢰 / 가격민감 — what kind of scrutiny this panelist is naturally suited to). Completion criterion: you can name the persona count and it is > 0.

2. Frame the target

State, in one line each: what the artifact is, and which question the panel answers. If the caller gave only "review this", use the default question. If the target is a URL or file you can open, open it first so panelists react to real content, not a guess.

3. Seat 5 panelists (spread, not convenience)

Pick five personas that spread across age bucket, region (province), and occupation — not five who cluster. A concrete rule that works: sort by a key that rotates each run (e.g. offset by how many times this target was reviewed), then walk the list skipping any persona whose (age_bucket, province) pair is already seated until you have five. If the question (step 2) maps to a review_lens — accessibility, pricing/trust, plain-language copy — prefer candidates whose lens matches when there's a tie, but never sacrifice the spread requirement to chase a lens match. Completion criterion: the five seated personas cover at least three distinct age buckets and four distinct provinces.

4. Dispatch the panel (parallel, independent)

Spawn five general-purpose agents in one message (five tool calls in a single turn, so they run concurrently). Give each ONE persona and the target. Panelist prompt template:

You are reacting to <target> only as this person — do not break character, do not review code or design systems, do not be agreeable to please. You are: <persona summary; age; sex; province; occupation; education; hobbies>. React honestly as this person would on first encounter:

1. In one sentence, what do you think this is?
2. What, if anything, confuses you or makes you hesitate?
3. Would you trust it / act on it? Why or why not?
4. One thing that would make it clearer or more convincing for you. Answer in Korean, in this person's voice and register. 6 sentences max.

Independence matters: do not let one panelist see another's answer. Do not answer on their behalf if a dispatch fails — re-dispatch or report the gap.

5. Synthesize

Collect the five reactions and write ONE report (schema below). Order it: reactions shared across panelists first (these are the strongest signals), then segment-specific friction (attributed to the persona segment that raised it), then prioritized recommendations. Route any user-voiced security/correctness worry to the owning reviewer under "Out of lane". Completion criterion: every finding traces to at least one seated panelist; no finding is invented to round out the report.

Output

## Persona panel review — <target>

**Panel** (5): <age·sex·region·occupation one-liners>

### Shared reactions
- <what most/all panelists felt> — <which segments>

### Segment-specific friction
- [<age·region·occupation>] <what tripped this persona> — <why it matters>

### Recommendations (prioritized)
1. <change> — addresses <friction>, for <segment>

### Out of lane (routed, not judged)
- <user-voiced security/correctness worry> → security-reviewer / code-reviewer

### Panel verdict
<one line: ships-for-users / needs-work / confusing — for whom>

Boundaries

• Not code/security/performance review — route those concerns; don't judge them.
• Not real user testing — a diverse, fast stand-in that surfaces likely friction early; say so when the stakes call for real users.
• Personas are synthetic — every persona is NVIDIA Nemotron synthetic data, grounded in real distributions but describing no real individual (any Korean name in a summary is generated, not a real person's). Don't treat a persona as a real person or add fields to it.
• No fabricated reactions — an unraised concern is not a finding.

Regenerating the catalog

The shipped catalog (119 personas) is a stratified subsample of the public nvidia/Nemotron-Personas-Korea dataset (CC BY 4.0): age_group × province hard-balanced across a 7×17 grid (one persona per cell — every province gets an even seat, not a population-proportional one, so a "spread panel" is always possible), with occupation_group × education_level soft-balanced on top. To rebuild it (e.g. a re-seeded sample), via DuckDB httpfs (no local dataset download):

uv run --with duckdb skills/persona-review/scripts/build_catalog.py --seed my-seed-2027

Attribution and modification notes live in the catalog's _meta block; keep them intact — CC BY 4.0 requires crediting the source and marking changes.