hook-mining

# Hook mining Most AI hook writing is slop for one of two reasons: it copy-pastes a hook that worked for someone else, or it "rewrites" it until the English is flat and the psychology is gone. This does neither. It keeps the **data for what worked** — the structure, the psychological mechanism, the rhythm — and swaps only the **power words**. Proven skeleton stays, skin changes. ## 1. Get real hooks in You cannot mine what you do not have. The input is a CSV of hooks that **actually performed**: - a **Sandcastles** (sandcastles.ai) export — it analyses the top videos in your niche and ranks the best hooks; export that tab to CSV - any social-analytics export with a hook/title/caption column and a views/likes/score column - competitor hooks pulled with the **agent-reach** skill, or breakdowns from **reel-analyzer** Column names are never predictable, so don't hand-parse. ## 2. Mine it ```bash python skills/hook-mining/mine_hooks.py hooks.csv [--top 40] ``` Auto-detects the hook column and the performance column, ranks, buckets by pattern, and flags the power words already present. Writes `hooks.mined.json`. Run `--selftest` to verify the parsing. Patterns it sorts into: `result-proof` · `time-bound` · `contrarian` · `secret-gap` · `listicle` · `how-to` · `warning-fear` · `newsjack` · `comparison` · `question` · `curiosity-tease` · `plain-statement`. ## 3. Remix — the part that matters For each hook worth reusing, produce 3–5 alternatives under one rule: > **Keep...