AI Persona Panels for Customer Research

Build reusable persona panels, ask open-ended customer questions, and run concept tests on messages, offers, and product ideas.

3-Step Workflow

Inspired by TinyTroupe (Generate Personas → Simulate Interactions → Extract & Analyze):

Recommended workflow:

1. /persona generate — build a reusable panel
2. /persona ask — explore motivations, barriers, language, and decision criteria
3. /persona concept-test — compare explicit options when ready for a structured choice task

Quality Guidelines

• Take your time with persona generation. Diverse, detailed personas are the foundation of useful results. Do not skip diversity verification.
• Do not skip validation. Always verify response JSON structure before proceeding to analysis.
• Quality over speed. A 5-persona panel with rich, differentiated responses is more valuable than a 15-persona panel with generic answers.

Quick Reference

ask explores open-ended motivations, barriers, and language. concept-test covers any comparison research: product concepts, messaging A/B, packaging, competitive comparisons, feature bundles, and value framing with price context.

For detailed input requirements and output descriptions per command, see references/command-details.md.

Argument Parsing Rules

The text after the command is parsed as follows:

1. Command (required): concept-test, generate, ask
2. Options (optional):

- --count N — Panel size (default: 5; not applicable for ask) - --market MARKET — Target market/country (default: us). Accepts country codes (us, japan/jp, uk, de, fr, cn, kr) or freeform descriptors ("Southeast Asia", "urban Brazil"). Shorthand codes are expanded: us → United States, japan/jp → Japan, uk → United Kingdom, etc. - --segments — Activate segment-driven flow; default panel size becomes 15 (not applicable for ask) - --panel PANEL — (ask only) Specify the panel survey-id to use (e.g., running-footwear-us-15p-2026-04). Skips auto-detection.

1. Free text (optional): Everything else is the research intent — a natural language

description of what the user wants to research. For ask, the free text is the question itself.

Parsing examples:

/persona concept-test Evaluate 3 new canned coffee concepts
  → type: concept-test, count: 5, intent: "Evaluate 3 new canned coffee concepts"

/persona concept-test --count 8 Compare ad headlines for EV launch
  → type: concept-test, count: 8, intent: "Compare ad headlines for EV launch"

/persona concept-test --segments Canned coffee
  → type: concept-test, count: 15, mode: segment-driven, intent: "Canned coffee"

/persona generate --count 15 Running shoes
  → type: generate, count: 15, market: "United States", intent: "Running shoes"

/persona generate --market japan --count 10 Canned coffee
  → type: generate, count: 10, market: "Japan", intent: "Canned coffee"

/persona concept-test --market japan Evaluate 3 new canned coffee concepts
  → type: concept-test, count: 5, market: "Japan", intent: "Evaluate 3 new canned coffee concepts"

/persona ask What frustrates you most about buying running shoes online?
  → type: ask, question: "What frustrates you most about buying running shoes online?", panel: auto-detect

/persona ask --panel running-footwear-us-15p-2026-04 Why would you ignore this ad?
  → type: ask, question: "Why would you ignore this ad?", panel: "running-footwear-us-15p-2026-04"

/persona ask --market japan What makes this product feel overpriced?
  → type: ask, question: "What makes this product feel overpriced?", market: "Japan", panel: auto-detect

When free text is provided:

• Extract the topic from the intent text
• Determine what information is still missing:

- concept-test: concept/option details (names + descriptions) — ask only if not in intent - ask: the question is the free text; no additional info required

• If sufficient information is present, proceed without further questions

When no free text is provided:

• Collect all required information interactively (current behavior)

Orchestration Logic

Step 1: Build Panel

1. Parse the user's request to determine:

- Command (concept-test or generate) - Free-text research intent (if provided) - Panel size (default: 5; or 15 if --segments) - Market (default: United States; resolve shorthands like jp → Japan) - Mode: topic-only (default) or segment-driven (--segments)

1. Collect missing info:

- Extract topic from free text (e.g., "Evaluate 3 new canned coffee concepts" → topic = "Canned coffee") - For concept-test: identify concept/option details — ask only if not provided - If free text provides sufficient context, skip interactive questions

1. Persona Panel Decision:

a) User explicit instruction: - "Use existing panel" / "reuse {name}" → load from specified personas/{survey-id}/ directory - "Generate new personas" or provides segment definitions → new generation - No instruction → proceed to auto-decision (b)

b) Auto-decision (default is NEW generation): - Scan personas/ for subdirectories containing manifest.json (ignore _archive/) - If a manifest's category closely matches the current topic AND market matches (or is absent, treated as "United States") → propose reuse, ask user - If no match found → proceed directly to new generation (no confirmation needed)

c) New generation flow — two modes:

Topic-only mode (default): - Skip segment inference entirely - Infer diversity dimensions inline and generate all N personas in one batch - See references/topic-only-generation-flow.md for the 4-step flow (dimension inference, target assignment, generation, manifest creation) - Do NOT ask user for segment approval — proceed directly

Segment-driven mode (--segments): - If user provided segment definitions → use them directly - If topic only → run segment inference (see references/segment-inference-prompt.md) to generate 3-4 segments → present to user for approval - Default panel size: 15 personas (3 segments × 5 personas) - Follow the 4-step generation flow below

### Segment-Driven 4-Step Generation Flow

Inspired by TinyTroupe's TinyPersonFactory plan-based approach: instead of letting the LLM freely generate N personas per segment, first build a deterministic sampling plan, then generate each persona from its slot specification.

Step 1a: Deterministic Count Allocation

Compute per-segment counts deterministically — do NOT delegate count management to the LLM. This is the orchestrator's responsibility:

1. base = floor(count / numsegments) 2. remainder = count mod numsegments 3. First remainder segments get base + 1; remaining get base 4. Verify: sum(segment_counts) == count before proceeding

Examples: - 30 ÷ 3 = 10 + 10 + 10 - 17 ÷ 4 = 5 + 4 + 4 + 4 - 8 ÷ 3 = 3 + 3 + 2

Record each segment's allocated count. If the sum does not equal the requested count, fix the allocation before generating.

Step 1b: Sampling Plan Generation

Generate a sampling plan that assigns diversity attributes to each persona slot. See references/sampling-plan-prompt.md for the prompt.

The plan is a JSON array with exactly N rows (one per persona). Each row specifies: segment, agebucket, gender, occupationtier, geographytype, regionhint, categorystance, ethnicityhint.

After the LLM returns the plan, verify: - len(plan) == count - Per-segment row counts match the allocation from Step 1a - If verification fails, regenerate the plan (max 2 attempts)

Step 1c: Persona Generation from Slot Specifications

Generate personas sequentially by segment. For each segment batch:

1. Pass all slot rows for that segment to persona-generation-prompt.md as {{slotspec}} — the LLM fleshes out each slot into a full persona 2. Pass the exclusion list of already-used names, surnames, and occupation titles from previous segments as {{exclusionlist}} 3. After generation, extract the names, surnames, and occupation titles from the generated personas and add them to the exclusion list for the next segment

This ensures cross-segment duplicate suppression is built into the generation flow, not left to chance.

Step 1d: Panel Validation

After all personas are generated, run scripts/validate_panel.py programmatically:

``bash python scripts/validate_panel.py --panel-dir personas/{survey-id} --requested-count {count} --json ``

Interpret the results: - Hard fails (severity: "fail"): The panel has a contract violation. Attempt to fix the specific issue (e.g., regenerate a duplicate-named persona) and re-validate. Max 2 retry cycles. - Warnings (severity: "warning"): Note them but do not block. - If hard fails persist after retries, present the panel with warnings and let the user decide.

d) Panel confirmation with QA summary: - Present the panel overview table: ```

`` - Below the table, show a QA Summary from validate_panel results: ` ### Panel QA - Count: 30/30 ✓ - Segment balance: Serious (10), Casual (10), Fashion (10) ✓ - Names: all unique ✓ - Occupation duplicates: none ✓ - Age spread: OK ✓ - Gender balance: OK ✓ - Big Five near-duplicates: 1 pair (0.985) ⚠ `` - If any hard fail remains, flag it prominently and offer to regenerate the affected personas - After user approves → proceed to concept test execution

e) Save personas: - Save each persona as personas/{survey-id}/{NameUnderscored}.json - Create manifest.json with: - "generationmode": "topic-only" or "segment-driven" - "market": "{resolved market name}" - "requestedcount": {count} - "samplingplan": [...] (the slot specifications from Step 1b) - "validationreport": {...} (from validatepanel output) - All segments must include "description" field

1. Load survey template from templates/concept_test.md.

Step 2: Run Concept Test

Read references/simulation-prompt.md for the core simulation prompt.

Before simulation, extract simulation profiles from the full persona JSONs using extractsimulationprofile() in scripts/simulatesurvey.py. Pass the compact profiles (~40-50 lines/persona) as {PERSONASJSON} instead of the full personas (~300 lines/persona). This saves context window while preserving all response-driving fields.

Agent-Separated Execution (Recommended)

Use simulate_survey.py for production-quality results. This wrapper must force --backend claude-cli, which runs each persona in a fully independent claude -p subprocess — zero inter-persona bias.

python scripts/simulate_survey.py --config {config-path} --backend claude-cli

How it works:

1. Builds a per-persona system prompt (simulation instructions + single persona profile + survey questions)
2. Launches parallel claude -p processes (controlled by --concurrency, default 5)
3. Each process uses --output-format json --tools "" --no-session-persistence, plus

--safe-mode (context isolation: no CLAUDE.md/plugins/hooks leak into persona calls) and --json-schema (server-validated structured output) when the installed CLI supports them

1. Parses the response (structured_output preferred, text extraction as fallback),

validates structure, retries on failure (up to 3 attempts)

1. Saves results.json + runmetadata.json to config.outputdir when provided, otherwise to outputs/{YYYY-MM-DD}/{HHMMSS}/{survey_type}/
2. Runs a backend preflight before persona fan-out; if it fails, stop immediately and record failure_stage: "preflight"

Options: --dry-run, --analyze, --report-llm, --no-adherence-check, --model, --concurrency, --report-backend, --no-structured-output, --no-isolation, --fallback-model, --effort

Model Selection

The default simulation model is sonnet — the best cost/quality balance for 10-15-way persona fan-outs. All claude model aliases and full model IDs work:

Config keys (all optional, omit for current behavior):

{
  "model": "sonnet",
  "report_model": "fable",
  "fallback_model": "sonnet",
  "effort": "medium",
  "max_budget_usd_per_call": 0.50,
  "structured_output": true,
  "isolation": true
}

• "report_model": "fable" — recommended upgrade: keep the fan-out on sonnet but

synthesize the final report with Fable 5 (one call, small cost increase, richer narrative).

• "fallback_model": "sonnet" — recommended when running the fan-out on fable/opus:

automatically falls back when the primary model is overloaded.

• "effort" — lower (low/medium) for cheaper, faster persona answers; not supported by haiku.
• "maxbudgetusdpercall" — hard cost cap per subprocess call.
• Run metadata records both the requested alias (resolved_model) and the exact

model IDs that actually served the calls (actualmodelids), plus totalcostusd.

Why agent-separated: Shared-context simulation suffers from anchoring bias, consensus bias, and style contamination. Agent separation eliminates all three. Context isolation (--safe-mode) extends this: persona subprocesses do not read the user's CLAUDE.md, plugins, or hooks, so project context cannot bias responses.

Error handling: If simulatesurvey.py exits non-zero, read runmetadata.json for failurestage, preflight.error, perpersona[].error, and perpersona[].validationissues. Common stages: preflight (backend unreachable), simulation (persona subprocess failed), validation (response parsing failed). Report the failure stage and the most relevant error payload before suggesting remediation.

The canonical persisted output format is:

[
  {
    "name": "Persona Name",
    "segment": "Profile Label or Segment Name",
    "age": 35,
    "gender": "Female",
    "occupation": "Nurse",
    "responses": {
      "question_key": "answer"
    }
  }
]

Step 3: Review Findings

1. Save canonical results.json + run_metadata.json
2. Run scripts/analyze_results.py immediately with --analyze
3. Generate report.md plus results.csv / summary.json and optional charts
4. Save to config.outputdir when provided, otherwise to outputs/{YYYY-MM-DD}/{HHMMSS}/{surveytype}/
5. Present a concise summary to the user with the report path
6. When this Claude wrapper requests an LLM report, keep --report-backend same so report generation also stays on claude-cli

Report generation defaults:

• ask and concept-test default to LLM-generated report.md when --analyze runs on a

claude-cli backend. Each survey type uses a dedicated synthesis prompt (ASKREPORTSYSTEMPROMPT / CONCEPTTESTREPORTSYSTEMPROMPT in scripts/analyzeresults.py).

• Pass --no-report-llm to force the rule-based template report (faster / cheaper, but less rich).
• Other survey types (brand-map, price-test, usage-habits, survey) still default to the

rule-based template; add --report-llm to opt into LLM synthesis.

Report format: See references/report-template.md for the baseline one-pager structure. The survey-type-specific LLM prompts embed their own richer structures (Direct Answer, Where They Agreed/Differed, Top Signals with clustering for ask; Preference Verdict, Segment Splits, Per-Concept Strengths/Weaknesses, Improvement Theme Clusters for concept-test).

Step: Run Ask (for /persona ask command)

/persona ask is a focused v1 addition that sits between generate and concept-test. It lets users pose open-ended qualitative questions to an existing panel — no new panel generation required. Each persona responds independently (agent-separated), and the answers are synthesized into a usable research summary.

1. Parse Command

Extract:

• Question (required): Everything after /persona ask [options] is the research question
• Panel override (optional): --panel <survey-id> — skip auto-detection
• Market (optional): --market <market> — used during auto-detection as a filter

If the user provides no question text, ask: "What would you like to ask the panel?"

2. Resolve Panel (required for ask)

/persona ask always requires an existing panel. It never generates new personas.

If --panel is given:

• Load personas/{PANEL}/manifest.json
• If the directory does not exist, stop and report the error

If no --panel given (auto-detect):

• Scan personas/ for subdirectories containing manifest.json (exclude _archive/)
• If --market is given, filter candidates to those whose manifest.market

matches (or whose market is absent, treated as "United States"). Drop non-matching panels.

• From remaining candidates:

- If exactly one candidate: use it directly — no confirmation needed - If multiple candidates: present a numbered list with survey-id, category, market, and persona count, then ask the user to pick one - If no candidates: tell the user clearly and preserve their question: > "No persona panel found. Run /persona generate {topic} first, then re-run your ask." Do NOT silently generate a new panel.

Display the matched panel to the user before running:

Using panel: running-footwear-us-15p-2026-04 (15 personas, US)

3. Build Ask Config

Construct a config JSON and save it to outputs/{YYYY-MM-DD}/{HHMMSS}/ask/ask-config.json:

{
  "survey_type": "ask",
  "panel_dir": "personas/{matched-panel-id}",
  "topic": "{manifest.category}",
  "variables": {
    "category": "{manifest.category}",
    "user_question": "{user's question verbatim}"
  },
  "backend": "claude-cli",
  "model": "sonnet",
  "max_concurrency": 5
}

4. Run Simulation

python scripts/simulate_survey.py --config {config-path} --analyze

The --analyze flag automatically runs analyzeresults.py --survey-type ask, which produces summary.json with pre-computed topsignals, emotion_distribution, plus report.md with ask-specific sections.

Each persona receives the question via templates/ask.md and returns:

• short_answer — direct 2-3 sentence answer
• reasoning — concrete personal reasoning
• themes — 1-4 short theme phrases
• emotion — one-word dominant tone

5. Present Ask Synthesis

After the run completes, an LLM-generated report.md already exists in the output directory (produced by analyzeresults.py using ASKREPORTSYSTEMPROMPT). Your job is to read that report plus results.json / summary.json, then:

1. Verify report.md quality. Check that it follows the expected structure:

Direct Answer → Key Findings → Where They Agreed → Where They Differed → Notable Verbatims → Top Signals → Emotion Distribution → Caveats. If any section is missing or weak, overwrite report.md with your own synthesis using the template below.

1. Present a concise summary to the user — not the full report, just 5-8 lines

covering the Direct Answer and top 2-3 findings. Always include the path to report.md so they can read the full document.

Use the pre-computed theme counts and emotion distribution from summary.json — do not recompute them manually. When presenting the summary to the user, follow this structure:

## "{question}" — Panel Response Summary
**Panel**: {manifest_topic} ({N} personas, {market})
**Date**: {date}

### Direct Answer
[2-3 sentence synthesis of what the panel said overall, without attributing to
any single persona. Focus on the dominant response pattern.]

### Key Findings
1. [Primary theme or point of agreement]
2. [Secondary finding or notable split]
3. [Surprise, tension, or segment difference]

### Where They Agreed
[What most or all personas said in common]

### Where They Differed
[Notable splits — by age, segment, personality type, or circumstance]

### Notable Verbatims
> "[direct quote from reasoning field]" — Name, Age, Occupation

> "[direct quote]" — Name, Age, Occupation

### Top Signals
[comma-separated theme phrases with frequency counts, e.g.: "sizing uncertainty (8)", "brand trust (5)"]

---
Artifacts: outputs/{date}/{time}/ask/
*Virtual panel of {N} synthetic personas — directional signal only.*

Important: This is a qualitative synthesis, not a concept test. Do NOT frame findings as preference shares or percentages unless emotion counts or theme counts are clearly labeled as such.

Step: Present Concept-Test Synthesis (for /persona concept-test)

Concept-test runs produce a structured choice (A/B/C) plus reasoning, purchase-likelihood scores, and improvement suggestions. After --analyze finishes, an LLM-generated report.md already exists in the output directory (produced via CONCEPTTESTREPORTSYSTEMPROMPT). Your job as the orchestrator is to:

1. Verify report.md quality. Check that it follows this structure:

Preference Verdict → Key Findings → Segment / Profile Splits → Purchase-Intent Drivers → Per-Concept Strengths and Weaknesses → Improvement Theme Clusters → Notable Verbatims → Caveats. If any section is missing or weak, overwrite report.md with your own synthesis using the template below.

1. Present a concise summary to the user — the verdict plus the 2-3 most actionable

findings, not the full report. Always include the path to report.md.

Synthesis template to use when overwriting report.md:

# Concept Test: {topic} — Virtual Research Report

**Date**: {date} | **Panel**: {N} personas

## Panel Overview
| # | Name | Age | Occupation | Profile | Preferred | Purchase Likelihood |
|---|------|-----|------------|---------|-----------|---------------------|
| 1 | ... | ... | ... | ... | ... | ... |

## Preference Verdict
[2-3 sentences. Leader + margin + confidence. Use "Clear winner" (>60%), "Narrow lead" (40-60%),
"No winner; tied" (ties), or "Fragmented" (every concept <40%). Include raw counts.]

## Key Findings
1. [Cross-persona insight]
2. [Second insight]
3. [Third insight]

## Segment / Profile Splits
[Who picked what. Organize by segment → preference → shared reason. 3-6 lines.]

## Purchase-Intent Drivers
[What moved `purchase_likelihood` up or down across the panel. Cite factors personas credited.]

## Per-Concept Strengths and Weaknesses
### Concept A — N/total chose this
**Strengths (cross-persona):** [synthesis, not per-chooser dump]
**Weaknesses (cross-persona):** [what passers flagged]

### Concept B — N/total chose this
[...]

## Improvement Theme Clusters
- **Cluster label** (N mentions across Concepts X, Y) — one-line gloss with 1-2 persona attributions
[3-6 clusters]

## Notable Verbatims
> "full sentence from reasoning field" — Name, Age, Occupation

> "another full sentence" — Name, Age, Occupation

## Caveats
- Virtual panel of {N} personas; directional only
- Not statistically representative; use for hypothesis generation
- AI-generated responses may exhibit positivity bias

Important: Synthesis over summary. Never write a "Profile Analysis" section with one paragraph per persona — that is the old rule-based format. Never dump per-chooser verbatim reasoning under a "Support reasons" list; always cluster across personas.

Persona Schema

See references/persona-schema.md for the complete JSON schema.

Key principle: Personas must be detailed enough to produce differentiated, realistic survey responses. The Big Five personality traits, communication style, and lifestyle details are what drive response variation.

In topic-only mode, the segment field contains a persona-specific archetype label (e.g., "Budget Pragmatist"). In segment-driven mode, it contains the shared segment name (e.g., "Daily Ritual Drinker").

Simulation Engine

See references/simulation-prompt.md for the core prompt template.

Key principles:

• Each persona's response must be deeply grounded in their profile
• Not all personas are positive or articulate — reflect real consumer diversity
• Big Five traits influence enthusiasm, skepticism, detail level, and decision style
• Education level affects vocabulary and expression complexity
• Some personas should give unexpected responses (break stereotype expectations)

Survey Templates

The concept test template (templates/concept_test.md) handles A/B/C concept evaluation with reasoning. It is flexible enough to cover product comparisons, message A/B tests, packaging evaluations, competitive comparisons, and feature bundle comparisons. For research that doesn't fit the template, the orchestrator constructs custom questions from the user's brief.

The ask template (templates/ask.md) handles open-ended qualitative questions. It elicits a direct answer, concrete reasoning, recurring signals, and emotional tone — structured enough for synthesis without forcing a choice.

Analysis Pipeline

Default (topic-only): Claude reads results.json and generates report.md directly.

--analyze flag: Runs scripts/analyze_results.py for:

• JSON → CSV conversion
• Cross-tabulation (segment × response) or persona comparison table
• Chart generation (bar, stacked bar, heatmap)
• Summary statistics
• Markdown report generation (--report-only skips charts)

Troubleshooting

File Organization

Skill directory (installed to ~/.claude/skills/persona/): references/, scripts/, templates/, demo/ — read-only assets, never modified during execution.

User's project directory (CWD): personas/ (generated panels), outputs/ (results, reports, charts). All generated data is saved to the user's current working directory, not the skill installation directory.

Full annotated tree in references/directory-structure.md.