market-simulation 0.9.3

AI market research simulation using NVIDIA Nemotron personas and Claude Code Agents

market-simulation

AI persona simulation via Claude Code — no local LLM or API key required.

Runs batched market research, copy testing, content audits, and more against
NVIDIA Nemotron-Personas (7 countries, 8M+ personas, CC BY 4.0).

⚠️ This is LLM-playing-LLM-personas. Results are hypotheses for early-stage validation,
not a substitute for real consumer research. Agreement rates skew high due to LLM positive bias.

---

Quick start

Requirements: Claude Code · Python 3.10+

pip install market-simulation
market-simulation install-skill

Restart your Claude Code session — the skill loads automatically.

---

Report preview

Dark-mode HTML report — sentiment distribution, demographic cross-tabs, auto-generated insights, and full response cards.

---

Example prompts

Use case	Prompt
Market reaction	"Simulate how 20 Seoul office workers in their 30s react to a ₩9,900/mo coffee subscription"
Copy A/B test	"Which of these 2 taglines resonates more with women in their 20s–30s in Tokyo?"
Content clarity	"Would a high-school-educated 40-year-old find this terms-of-service hard to understand?"
Chatbot tone check	"Does this chatbot response feel natural to male users in their 50s?"
Policy / HR	"Compare reactions to a 4.5-day workweek across manufacturing, IT, and service workers"
Brand naming	"Which of these 3 brand names feels most trustworthy to self-employed people in their 30s?"

---

Supported countries

country	Dataset	Language
korea	Nemotron-Personas-Korea	Korean
usa	Nemotron-Personas-USA	English
japan	Nemotron-Personas-Japan	Japanese
india	Nemotron-Personas-India	English / Hindi
france	Nemotron-Personas-France	French
brazil	Nemotron-Personas-Brazil	Portuguese
singapore	Nemotron-Personas-Singapore	English

---

How it works

HuggingFace streaming       Claude Code skill
(8M+ personas)   ──▶   filter target segment
                 ──▶   split into batches of 5
                 ──▶   run parallel sub-agents (isolated context per batch)
                 ──▶   collect responses → CSV + HTML report

• Streaming load — no full dataset download required
• Isolated batches — no cross-contamination between personas
• Output: output/YYYY-MM-DD_{topic}.csv + .report.html

---

Simulation limits

	Value
Default	20 personas
Maximum	30 (6 agents × 5)

---

Programmatic use

from market_simulation import load_pool, filter_pool, occupation_kw

df = load_pool('usa', sample_n=50000)
pool = filter_pool(df, province='CA', age_range=(25, 39),
                   occupation_keywords=occupation_kw('tech'))
sample = pool.sample(20, random_state=42)

English occupation filters: occupation_kw('tech'), occupation_kw('finance'), occupation_kw('healthcare'), etc.

---

Disclaimer

• Results are LLM-generated hypotheses — not survey or interview data.
• Use relative comparisons within a run. Absolute numbers (e.g. "65% positive") are inflated by LLM positive bias.
• Persona data is fully synthetic. Any resemblance to real individuals is coincidental.
• Enterprise-oriented personas (finance, healthcare, etc.) are not included in the dataset. Filters for those occupations may return thin pools.
• Persona data: CC BY 4.0 (NVIDIA). Attribution required when publishing results.

---

License

Code: MIT · Persona data: CC BY 4.0 (NVIDIA)