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
bashpython 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 the psychology. Swap the power words.
Hold fixed: the pattern, the clause order, the specificity (a number stays a number), the rhythm and syllable shape, the tension and where it resolves, the point of view.
Swap only: the charged words — the verb, the intensifier, the noun carrying the emotion.
power_word_frequency in the mined JSON tells you which ones are earning their keep in your niche.
Pull replacements from there, not from a thesaurus.
proven: "You're using Claude wrong" [contrarian · you're + wrong] ✅ "You're prompting Claude backwards" swapped the verb + the charge word ✅ "You're building agents the slow way" same skeleton, same tension ❌ "Many people use Claude incorrectly" psychology gone, flat English → slop ❌ "You're using Claude wrong" verbatim copy → plagiarism, the platform buries it
Then hand the winners to viral-hook-writer to rank, or straight into reel-scripter.
Hard rules
- Never ship a mined hook verbatim. It is someone else's line. Instagram's originality systems bury copies, and it's plagiarism. Model the technique, write your own words.
- Never remix the structure. Change the clause order or drop the number and you have thrown away the only thing the data proved.
- A number in the source stays a number in the remix — and it must be YOURS and real. Swapping in a stat you cannot defend is the fastest way to lose an audience.
- Rank before you mine. Remixing a hook that did not perform just launders a bad idea.
- Small buckets are signal. A pattern with two entries that both went viral beats one with forty mediocre ones. Weight by performance, not by count.
Failure modes
| Symptom | Cause | Fix |
|---|---|---|
| Variants read flat / corporate | rewrote the sentence instead of swapping words | restore the original skeleton, change only the charged words |
| All variants sound identical | swapped one word in one slot | vary the slot — verb, intensifier, or emotional noun |
| Variants lost the tension | resolved the curiosity gap inside the hook | the gap stays open; the payoff belongs after the hook |
| Nothing outperforms | mining unranked or off-niche data | re-export ranked, filter to the actual niche |
Where it sits
agent-reach / reel-analyzer find what's working → hook-mining turns that data into your own
hooks → viral-hook-writer ranks → reel-scripter writes the script → reel-builder builds it.
