Generate hundreds of AI personas with real behavioral psychology. Simulate them through your pricing page, your copy, your feature set. Watch them reason through the decision.
The verdict isn't the point. The reasoning is.
Persona #2 of 128. A 58-year-old school administrator in Omaha encounters a $49/year subscription. Her first instinct isn't to evaluate the price — it's to check the refund terms. The reasoning traces back to a home warranty claim denied on a technicality in 2011.
You describe a product. Facet generates a population of psychologically detailed personas — each with a specific income, decision style, existing subscriptions, and a reason they might say no. Then it runs each one through your pricing, copy, or feature decisions.
Each persona gets:
- Behavioral economics profile — reference point, loss aversion, mental accounting, subscription fatigue, status quo bias
- Decision simulation — a four-step emotional arc (gut reaction → building considerations → moment of doubt → resolution)
- Verdict — beliefs, desires, intentions, and a representative quote in their own voice
The synthesis across all personas includes confidence grading, a sycophancy audit, a pre-mortem failure narrative, and a validation plan for what to test with real users.
Facet generates hypotheses, not evidence. The output looks like a $50K research engagement — rich narratives, specific numbers, emotional arcs — but it's all generated by an LLM that has never met a real user.
Good for: Thinking through product bets before you commit resources. Identifying objections you hadn't considered. Stress-testing pricing models and copy variants. Preparing sharper questions for real user interviews.
Not good for: Final pricing decisions. Go/no-go launch calls. Feature priority rankings (LLMs care about everything equally). Non-Western or underrepresented populations.
The accuracy hierarchy is clear:
| Approach | Accuracy | Example |
|---|---|---|
| Interview-grounded agents | ~85% | Stanford, 2024 (founding team of Simile) |
| Fine-tuned on survey data | ~78-88% | Qualtrics Edge Audiences |
| Census-calibrated panels | ~75-92% | Ditto, Toluna |
| Facet + your data | ~78-85% | --calibration with internal research |
| Rich prompting | ~70-75% | Facet (default) |
| Naive LLM query | ~63% | "Hey ChatGPT, would you buy this?" |
Facet is in the fifth row by default. Feed it your own research data — past surveys, support tickets, interview transcripts, analytics — and it moves to the fourth. The Stanford research that achieved 85% did so by grounding agents in real interview data. Facet's --calibration flag does the same thing.
Free, open source, and auditable.
git clone https://github.com/saraswatayu/facet.git
cd facet
./setup
# 1. Generate personas for your product
./sim.sh init --config examples/superhuman-product.md --name superhuman
# 2. Run studies against those personas
./sim.sh study --panel output/superhuman/ --config examples/superhuman-pricing.md
./sim.sh study --panel output/superhuman/ --config examples/superhuman-copy.md
# 3. Check status
./sim.sh status --panel output/superhuman/Personas are generated once and reused across studies. A pricing study and a copy study share the same persona backgrounds.
| Flag | Description | Default |
|---|---|---|
--config |
Product config (init) or study config (study) | required |
--name |
Study name for output directory | config filename |
--panel |
Path to existing panel directory | — |
--runs |
Number of simulation runs for stability testing | 1 |
--concurrency |
Parallel persona generations/simulations | 5 |
--calibration |
Real research data to ground personas (details) | — |
--output-dir |
Override base output directory | ./output/ |
If you use Claude Code, install Facet as a plugin for a conversational research experience:
# As a skill (personal, available in all projects)
git clone https://github.com/saraswatayu/facet.git ~/.claude/skills/facet
# Or as a plugin (for development/testing)
git clone https://github.com/saraswatayu/facet.git
claude --plugin-dir ./facetThen ask a product question:
/facet "Should I charge $15 or $30/month for my task management app?"
The skill scans your codebase for pricing and feature data, asks 2-3 calibration questions, generates a study config, and runs the full simulation phase by phase with progress updates. When it finishes: a research panel introduction, findings with honest confidence levels, and follow-up options. You'll get a desktop notification when the study is ready.
You can also invoke the research agent directly: @facet-researcher run a pricing study for this product. The agent has persistent memory and remembers past studies across sessions.
Run automated research without interaction:
claude -p "Run a Facet pricing study for this product with options at \$15 and \$30/month" \
--allowedTools "Bash,Read,Write,Glob,Grep" \
--output-format jsonINIT — run once per product
┌──────────┐ ┌──────────────────────────────────┐
│ PLAN │ │ GENERATE (parallel, in waves) │
│ │─────▶│ │
│ segments │ │ 1 claude call per persona │
│ personas │ │ wave-based diversity context │
│ matrix │ │ behavioral economics profiles │
└──────────┘ └──────────────────────────────────┘
STUDY — repeat for pricing, copy, features, onboarding, retention
┌──────────────────────────┐ ┌──────────────────┐
│ SIMULATE (parallel) │ │ ANALYZE │
│ │─────▶│ │
│ Chain-of-Feeling arcs │ │ synthesis.md │
│ BDI verdicts │ │ artifacts.md │
│ deal-breaker enforcement│ │ counterargument │
└──────────────────────────┘ └──────────────────┘
Reads your product config. Generates segments based on jobs-to-be-done (not just demographics — research shows demographics explain only ~1.5% of response variance). Each persona gets a constraint vector:
Persona #7 — Segment: Replacing a Manual Workaround
Name: Marcus | Age: 41 | City: Boise, ID | Job: Operations Manager
Income range: $72-88k | Household: married, two kids
Big Five: O:5 C:8 E:4 A:3 N:5
Decision style: optimizer | Adoption stage: early majority
Status quo bias: strong | Processing mode: System 2
Current solution: custom spreadsheet he's maintained for 3 years
Deal-breakers: won't give any service access to his email account
One Claude call per persona, in parallel waves of 5. Each wave sees summaries of previously generated personas to prevent homogeneity. Anti-sycophancy rules at every phase — personas are explicitly permitted to reject the product.
Each persona walks through your pricing page / copy / features with a four-step emotional arc: gut reaction, building considerations, moment of doubt, resolution. Deal-breakers from the background are enforced — not every persona converts.
Reads all personas + simulations. Produces:
- synthesis.md — confidence-graded findings, behavioral mechanism analysis, stated vs. revealed preference gaps, sycophancy audit, pre-mortem, counterargument
- artifacts.md — usable copy, FAQ from objections, segment marketing angles, referral templates, validation plan
---
segments: 10
personas_per_segment: 5
---
# Product: Superhuman
Superhuman is a premium email client built for speed...
## Key Product Details
- $30/month per user, works with Gmail and Outlook
- AI features: summarize threads, draft replies, auto-triage
## Target Market
- Startup founders and CEOs
- Sales professionals---
study_name: pricing-tiers
study_type: pricing
options:
- name: "Model A"
description: "$30/month flat — current pricing"
- name: "Model B"
description: "Free tier + $30/month Pro"
copy_variants: true
---
## Options to Test
### Model A: $30/month flat (status quo)
One plan, one price, every feature...
## Copy Variants
### Variant A: "Premium, Unapologetic"
$30/month. The fastest email experience ever made.Study types: pricing, copy, features, onboarding, retention
Ground personas in real research data instead of relying solely on LLM priors:
# Single file
./sim.sh init --config product.md --name study --calibration survey-results.md
# Directory of research files
./sim.sh init --config product.md --name study --calibration ./research-data/For directories, add a manifest.md at the root describing each file's purpose. The plan phase reads relevant files and documents what it extracted in a "Calibration Sources" audit trail.
Templates are grounded in ~490 academic sources across six research areas. The short version:
- LLM personas are sycophantic. They agree with you. Templates include anti-sycophancy rules at every phase, explicit permission to reject, and adversarial review that tries to break the recommendation. (Sharma et al., ICLR 2024)
- LLM personas are homogeneous. 30-50% less variance than real humans. Wave-based generation with diversity context, constraint vectors, and a diversity matrix with concentration thresholds. (Bisbee et al., 2024)
- LLMs don't naturally produce human decision biases. Only 17-57% consistency. Explicit behavioral economics parameters per persona: loss aversion, mental accounting, subscription fatigue, status quo bias. (arXiv:2509.22856)
- Demographics explain almost nothing. 1.5% of response variance. Segments use jobs-to-be-done. Personas use values and behavioral profiles, not just age/income. (SCOPE framework, arXiv:2601.07110)
Full reports in research/.
output/{product}/
├── plan.md # segments, constraint vectors, diversity matrix
├── personas/ # generated once, reused
│ ├── persona-001.md
│ └── ...
└── studies/
└── {study-name}/
├── simulations/
│ ├── persona-001.md # decision arcs, verdicts
│ └── ...
├── synthesis.md # analysis + counterargument
└── artifacts.md # deliverables + validation plan
- Illusion of depth. "$38,500/year, $18.50/hour" is plausible fiction, not data. The richer the output, the easier it is to forget this.
- Can't know what it can't know. Real users have DIY workarounds and half-formed habits no LLM can simulate.
- Config bias. If your study config describes Option A more favorably, the entire simulation leans toward A.
- Run-to-run variance. Identical parameters can produce different results.
- WEIRD bias. LLMs are overtrained on English-language, Western, educated, middle-class perspectives.
- Claude CLI (
claudecommand available) - Python 3 (for
stream_filter.py) - Bash
Run ./setup to verify dependencies.
MIT