Creative Signal

Correlate creative attributes with hook rate (3-second views ÷ impressions) and hold rate (thruplay ÷ 3-second views) across every video ad that delivered in the window. Scripts pull metrics into a local SQLite store, tag each creative once (Gemini gemini-3.1-flash-lite + ffmpeg features), and compute per-attribute lift with explicit confidence. The agent reads signals.json and authors brief.md — judgment on top, never re-derived math.

Arguments: $ARGUMENTS — optional window (default last_14d; valid last_7d, last_14d, last_30d), or reconfigure, or backfill.

v1 makes no revenue or ROAS claim. Hook and hold are attention proxies. State this in every brief.

Data Architecture

<skill-dir>/                          # this folder — Base directory injected when the skill loads
├── .venv/                            # created by onboarding; every script runs via .venv/bin/python
├── scripts/                          # run.py orchestrates the rest
└── references/                       # onboarding.md, taxonomy.md, gemini-prompt.md