Yrzhe Social Sandbox
Thin scenario wrapper for counterfactual social simulation. Differs from product-feedback (scores products) and vote-predict (categorical votes) — this produces open-ended qualitative narratives from the panel, then clusters them.
Recipe
import sys
sys.path.insert(0, str(__import__('pathlib').Path.home() / '.claude/skills/persona-sim'))
from lib import sampler
from lib.sim_engine import SYSTEM_PROMPT, _persona_card
from lib.llm_router import generate
from concurrent.futures import ThreadPoolExecutor
# 1. Sample diverse panel (wider demographic spread than product-feedback)
panel = sampler.sample_personas(n=50, source="nemotron_usa", mode="stream")
# 2. Design the scenario — use past tense as if it already happened
scenario = """
Congress has just passed a law raising the federal minimum wage to $22/hour
nationwide, effective in 6 months. You've seen the news today.
"""
# 3. Ask each persona: immediate reaction + 6-month expectation + what they plan to do
def probe(persona):
task = (f"{scenario}\n\n"
"Respond as yourself, in 3 sentences:\n"
"(1) Your immediate emotional reaction.\n"
"(2) What you expect to happen in your life over 6 months.\n"
"(3) What (if anything) you plan to do in response.")
return generate(system=SYSTEM_PROMPT, persona_card=_persona_card(persona),
task=task, tier="default", max_tokens=400)
with ThreadPoolExecutor(max_workers=8) as ex:
narratives = list(ex.map(probe, panel))
# 4. Cluster the narratives (use Sonnet for this, not Haiku)
# Pass the 50 narratives + panel demographics to Sonnet:
# "Identify 3-5 distinct reaction archetypes. For each: label, % of panel, key demographic correlates,
# sample quote."
What this skill is for
- Hypothesis generation before designing a real survey
- Finding dimensions of disagreement you hadn't thought of
- Surfacing minority voices that demographic-only polling would miss
- Stress-testing messaging against diverse interpretations
What this skill is NOT for
- Predicting actual policy outcomes (LLM personas don't model emergence, network effects, or real economic constraints)
- Replacing real qualitative research (LLM can't replicate lived-experience nuance)
- High-stakes decisions — treat outputs as "interesting starting points" not evidence
Output structure
Always produce:
- 3-5 archetypes with % of panel and 1-2 sample quotes each
- Dimensional axes of disagreement (e.g., "rural vs urban", "service workers vs knowledge workers")
- Surprising/outlier reactions — these are often the most valuable signal
- Calibration warning from
bias_audit— note that LLM personas under-represent certain human biases
Cost note
n=50 × 1 open-ended call at Haiku + 1 Sonnet clustering call ≈ $0.50-1 per sandbox run.
See also
persona-sim/references/methodology.md— Park 2024 narrative-first finding is especially relevant herepersona-sim/lib/aggregator.py— useful for quantitative follow-ups on archetype proportions