product-feedback-sim

Product Feedback Sim

Thin scenario wrapper around persona-sim. Runs SGO on a product/copy candidate against a target user panel.

Recipe

import sys
sys.path.insert(0, str(__import__('pathlib').Path.home() / '.claude/skills/persona-sim'))
from lib import sampler, sim_engine

# 1. Define target audience
panel = sampler.sample_personas(
    n=30,                                              # 30-50 is cost-effective
    filters={"occupation_isco": "software", "age": (25, 50)},  # substring match on occupation
    source="nemotron_usa",
    mode="stream",
)

# 2. Single-version feedback (fast)
base = sim_engine.panel_score(
    panel,
    target="<your product/copy/feature description>",
    goal="<what you want to maximize, e.g. paid conversions>",
)
# → aggregate.histogram / median / iqr / disagreement_flag / by_age / by_gender

# 3. A/B (or N-way) gradient ranking — uses anchored counterfactuals on persuadable middle
ranked = sim_engine.sgo(
    panel,
    target="<original version>",
    candidates=["variant A", "variant B", "variant C"],
    goal="<same goal>",
)
# → ranking: [{"candidate", "avg_score_lift", "n_probed"}, ...]  sorted descending

Decision rules

• Use panel_score only when you want to understand a single version (no alternatives yet).
• Use sgo ONLY when there's a persuadable middle (score 4-7). If base median is ≤3 or ≥8, SGO has no signal — iterate on the base version first.
• n=30 minimum for stable distribution; n=50-100 if you need tight CIs.
• Always expose disagreement_flag to the user — high-variance results are signal, not noise.
• Always prepend the disclaimer from result["warning"] when showing output.

Target selection guidance

If the user says "我的产品用户是 X"：