korean-synthetic-consumer

name	korean-synthetic-consumer
description	Korean synthetic consumer/persona panel skill using Hugging Face nvidia/Nemotron-Personas-Korea plus lightweight report templates. Use for message-test, product-fit, objection-map, persona-panel, interview-script, full-report, setup.

Korean Synthetic Consumer / 한국형 가상 소비자 패널

Purpose: create a Korean synthetic consumer/persona panel grounded in Korean demographic synthetic data, then use the panel to pressure-test products, sales messages, proposals, and interview plans.

This skill is a portable version of Paul's synthetic-consumer workflow with dataset access added. It is especially useful for Phai/WAINBOX/ARMY ideas, proposal review, AX consulting, and Korean market message testing.

Safety caveat

Synthetic personas are for hypothesis generation. They are not proof of market demand and do not replace real customer interviews, surveys, sales calls, or behavioral data.

Recommended architecture

Claude/OpenClaw skill
  + Hugging Face/DuckDB sampler
  + markdown report templates

• The assistant uses this SKILL.md to choose the right mode and questions.
• scripts/sample_personas.py samples a small panel from nvidia/Nemotron-Personas-Korea or a local JSON file.
• scripts/build_panel_report.py prepares a structured prompt/report scaffold.
• DuckDB/parquet caching is optional. A shared team cache is preferred; every individual does not need to manually install or commit the full dataset.

Modes

message-test

Use for landing copy, KakaoTalk outreach, sales one-liners, proposal intro, Phai/WAINBOX/ARMY positioning.

product-fit

Use for concept fit, actual usage moment, adoption trigger, workflow friction, WTP.

objection-map

Use before sales/proposal review to map budget, trust, security, workflow, approval, and outcome objections.

persona-panel

Use for quick first reactions from 3-7 diverse personas.

interview-script

Turn synthetic signals into non-leading real customer interview questions.

full-report

Produce a full report: persona table, reaction distribution, quotes, WTP/adoption conditions, objections, recommendation, next validation.

setup

Install optional dependencies and verify scripts:

python -m pip install -r requirements.txt
python scripts/sample_personas.py --dataset examples/example-personas.json --n 2 --output markdown

Workflow

1. Clarify target: product/message, audience, business context, decision needed.
2. Sample personas:

   python scripts/sample_personas.py --n 5 --region 서울 --output json > personas.json
   

   Offline fallback:

   python scripts/sample_personas.py --dataset examples/example-personas.json --n 3 --output json
   

3. Build scaffold:

   python scripts/build_panel_report.py --personas personas.json --product "..." --mode full-report
   

4. Run the actual reasoning/interview synthesis in Claude/OpenClaw using the scaffold and relevant template.
5. End with: what to test with real customers next.

Anti-bias rules

• Include at least one skeptical persona.
• Avoid polite positivity; ask what feels overclaimed, risky, or not worth changing behavior for.
• Separate positive, conditional, and negative reactions.
• Treat WTP as directional only.
• Always state the next real-world validation step.

Validation ladder

Before treating outputs as business evidence, follow docs/validation-ladder.md: synthetic panel → real interview → behavioral test → decision memo.