Skip to content

Latest commit

 

History

1 Commit

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 

Repository files navigation

/fgi — Synthetic Focus Group Interview Moderator

A Claude Code skill that runs a qualitative Focus Group Interview (FGI) on your MVP, design mockup, or policy draft — using N synthetic personas simulated in isolated contexts, moderated through a disciplined 5-phase session, and distilled into a thematic report that preserves verbatim Voice-of-Customer quotes.

Unlike a single "act as a user and give feedback" prompt, /fgi reproduces the structure of real focus group research: balanced moderation, cross-persona reactions, group-effect dynamics, and an explicit honesty guard about what synthetic feedback can and cannot tell you.

⚠️ Honest by design. Synthetic personas are good for discovery, not validation. Every report ends with a "validate with 5 real users" checklist. This is a feature, not a disclaimer.

Why it's different

  • Persona isolation — each persona runs as a separate subagent, so they don't collapse into one averaged "AI voice" (a known failure mode of multi-persona prompting; see Nature Machine Intelligence 2025 on identity flattening).
  • Methodology-anchored — the 5-phase flow, balanced-intervention rules, and thematic analysis are drawn from focus-group research literature (Krueger 1998, Focus Agent arXiv 2409.01907), not invented.
  • Group dynamics — cross-persona probes ("anyone disagree?"), dominant-speaker balancing, and conflict surfacing are built into the moderator loop.
  • Verbatim preservation — themes are backed by raw persona quotes, never summarized away.

How it works

/fgi <domain> <stimulus>
   │
   ├─ 1. Recognize domain & stimulus (text or HTML mockup)
   ├─ 2. Recommend N personas (3 representative + 1 edge)   🚪 you approve
   ├─ 3. (optional) Pre-filter the stimulus for obvious weaknesses
   ├─ 4. Dispatch one isolated `fgi-participant` subagent per persona
   ├─ 5. Run the 5-phase session (priming → icebreak → general → stimulus deep-dive → wrap-up)
   ├─ 6. Thematic analysis (utterance → code → sub-theme → theme)
   └─ 7. Report (Executive Summary · Themes w/ verbatim VoC · Recommendations · Limitation guard)

The 5 phases

Phase Length Purpose
① Priming short rules, purpose, "please be honest"
② Icebreak short persona self-intro, natural entry
③ General medium domain opinions, no stimulus yet
④ Stimulus + deep-dive long expose stimulus + 3–5 probes + cross-persona reactions
⑤ Wrap-up short summary, live ranking, final messages

Research backing

Installation

Copy into your Claude Code skills directory:

cp -r fgi ~/.claude/skills/fgi

Restart Claude Code so the skill registers.

Companion subagent

/fgi dispatches an isolated subagent named fgi-participant (one per persona). You'll want a matching agent definition in ~/.claude/agents/fgi-participant.md whose job is: receive one persona's attributes + the stimulus + current moderator question, and respond in first person as that persona only. A starter definition is included in agents/ if present; otherwise any single-persona role agent works.

Usage

/fgi sleep ./mockups/sleep-onboarding.html
/fgi <your-domain> "a one-line description of your concept or mockup"

The skill ships with one example persona set in references/ (personas-sleep.md). Duplicate it and edit to build a persona set for your own domain — keep the single / cluster / conflict / edge group balance.

The personas

Each example set has 8 personas across 4 group roles:

  • single (S) — one dominant need (e.g. family-focused, value-focused)
  • cluster (M) — 2–3 mixed needs
  • conflict (C) — competing needs (e.g. family SUV and driving thrill)
  • edge (E) — extreme / outlier (e.g. low digital literacy, skeptic)

Each persona carries: age, occupation, lifestyle, expectations, concerns, tone of voice, and forbidden phrases (to keep the simulation in character).

⚠️ Limitations — read this

Synthetic FGI output is a discovery-grade hypothesis, not validated user research. Before any high-stakes decision (shipping a feature, finalizing a policy), validate with:

  • 5+ real-user IDIs or a mini-FGI
  • a usability test (Nielsen's 5-user rule)
  • quantitative metrics (conversion, dwell time)

License

MIT — see LICENSE.

Credits

Built by Sophia in Korea — a non-engineer who ships agents anyway. Methodology anchored in the focus-group research cited above; persona examples are fictional and illustrative.

About

Run a qualitative focus group on your MVP using N synthetic personas. A Claude Code skill — discovery-grade, honest about its limits.

Topics

Resources

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors