← ClaudeAtlas

hook-rubriclisted

How hook-scout scores candidate stream moments 0–10 for short-form potential, picks a rough window, and rejects moments that need outside context. Carries the scoring rubric, the window rules (15–60 s, payoff inside, first 2 s must hook), the overlap rule, and the allowed payoff tease. Use whenever candidate moments are scored or a clip window is chosen.
vaibhav-barman/live-clipper · ★ 0 · AI & Automation · score 72
Install: claude install-skill vaibhav-barman/live-clipper
# HOOK RUBRIC — is this moment a clip? > **THE JOB:** read ±30 s of transcript and signals around each candidate, score it, give a one-line > reason, propose a window. A score without a reason is rejected by the loop check. ## Inputs (from the ledger, never from video) - transcript words with times (voice stem), ±30 s around the candidate - signals: loudness spike (dB over the rolling median), chat messages/sec vs baseline, emote bursts, scene cuts, keyword hits - at most 3 still frames per candidate (only when the transcript alone is ambiguous) ## The score (0–10, sum of five parts) | part | 0 | 1 | 2 | |---|---|---|---| | **Hook** — do the first 2 s grab on their own? | slow build-up | interesting but slow | instant: a shout, a reveal, a question, a fail | | **Payoff** — is there a clear peak/punchline? | none | soft | unmistakable | | **Self-contained** — makes sense with zero context? | needs the last hour | needs one line of setup | stands alone | | **Reaction** — does the streamer/chat react? | flat | some | big (signals confirm it) | | **Shareability** — would someone send it to a friend? | no | maybe | yes | Keep only **≥ 7**. Below that, log it as `rejected` with the reason so the Feedback graph can learn. ## Before scoring - `get_page()` → the page's **niche angle**. Moments on-angle score as normal; off-angle moments need a strong reason in the line. - `get_learning()` → apply the **Current Rules** (they come from this page's own data). - `get_job` → th