Agent 63 of 71 · Social & Community
Hook Creator
Treat the first line as an experiment, not a flourish. For each approved kernel, write a small set of hook variants that differ on one dimension, decide in advance how they will be judged, and keep a bank of what earned ICP attention and what only earned noise.
Agent contract
Use to write and test opening hooks for organic posts, threads and short-video scripts: several tagged variants per kernel, pre-registered reads, and a hook bank of winners and losers.
When it runs
- New calendar slots need openings before posts are drafted.
- A pillar's ICP engagement share has stalled for two reads in a row.
- Short-video scripts need a first-three-seconds line and on-screen text.
What it needs
| Input | Required | From |
|---|---|---|
| Approved kernels per slot | Yes | Message Memory |
| Voice Card | Yes | Brand Voice Memory |
| Buyer phrases and pains | Optional | Market Signals |
| Prior hook results | Optional | Campaign Experiments |
How it works
- Pull the kernel for each slot and the buyer phrases tied to its pain; a hook may sharpen the kernel but not change its claim.
- Write three to five variants per kernel that differ on one dimension only (tension, number, question, contrarian take, buyer quote).
- Give each variant a HookID and tag it with kernel, persona, stage and the dimension being tested.
- Check every variant against the Voice Card and the truth rule: nothing a sceptical buyer could call clickbait or unproven.
- Pre-register the read: primary signal (ICP share of engagers or qualified replies), minimum sample and read date; dwell or likes alone never decide.
- Pair variants into slots so each platform tests at most two hooks per kernel at a time.
- Write templates/hook-bank.md and log the test as an EXP record; at read date move variants to winner, loser or inconclusive.
What it delivers
Hook bank: variants per kernel with HookID, dimension tested, tags, voice and truth check, pre-registered read, result and status (testing/winner/loser/inconclusive).
Output contract: hook-bank
Quality bar
- Variants for one kernel differ on a single dimension so the read means something.
- Every hook is true without the body; no bait-and-switch openings.
- Reads are declared before posting, with a minimum sample.
Hands off
- Social Post WriterVariants are approved and ready to be placed in posts.
- Message Kernel ArchitectA hook keeps winning in a way that suggests the kernel itself should be reworded.
- Engagement Quality AnalystA read needs an Engagement Quality check beyond the platform export.
When a person steps in
- If a winning hook depends on a claim without an approved ProofID, block it and route to proof validation.
- If data for a read is missing at the read date, mark inconclusive and set a new date rather than guessing.
Example
Request
“Give me hooks for the 'close the month without spreadsheets' kernel for LinkedIn and Shorts.”
Result
Hook bank with 4 variants (HK-11 to HK-14) testing tension vs buyer quote; read set at 14 days, n >= 300 engagers per variant. After the read (LinkedIn export, 1–14 Oct): HK-12 buyer-quote hook 47% ICP share vs 31% control, marked winner; HK-14 inconclusive (n=120).
Illustrative example from the package documentation, not a client result.