verify-bookends

name	verify-bookends
description	Post-edit verification of a short's opening and closing 1.5s. Claude sees a 3-frame strip from each end plus the trimmed transcript text around each end and returns {action:keep}, {action:trim,t0,t1} (inward-only), or {action:drop,reason}. Checks three things — (1) cleanliness (partial words, breath cutoffs, off-shot frames); (2) opening-hook strength (snap t0 inward to a stronger hook line if the first 3s is pure setup); (3) payoff landing (snap t1 to ~80ms past the payoff word when the tail trails into filler). Runs after tighten-pace, before fit-vertical. Inward-only — bookend-trim's outward pass already had its chance. Drops the span only on cleanliness failures requiring >2s of trim — never on hook-weakness alone.
allowed-tools	Bash
user-invocable	true

verify-bookends

Last-line defense before a short is committed to fit-vertical. Checks three things in one Claude call:

1. Cleanliness — partial words, breath cutoffs, co-speaker interjections, off-shot frames at either end.
2. Opening-hook strength (VVSA gate) — the first ~3s is where a scrolling stranger swipes away. If the opening is pure setup / throat-clearing and a stronger hook line exists within the first 3s, snap t0 inward to that line. Hook-weakness alone is never a drop — only an inward-trim suggestion.
3. Payoff landing — if the tail contains a clear payoff word followed by trailing filler ("yeah", "so anyway") or silence, snap t1 to ~80ms past the payoff word rather than the sentence boundary. Dead air after the payoff costs retention.

Output remains a single keep/trim/drop decision — same I/O contract, same caller logic in shorts.sh.

Invoke

.claude/skills/verify-bookends/verify-bookends.sh <input_clip> <input_trimmed_transcript> <out_decision_json> [VERIFY_BOOKENDS=0 to skip]

• input_clip — post tighten-pace clip
• input_trimmed_transcript — post tighten-pace re-timed transcript
• out_decision_json — written JSON: {"action":"keep"}, {"action":"trim","t0":<sec>,"t1":<sec>,"reason":...}, or {"action":"drop","reason":...}

Env: VERIFY_BOOKENDS=0 makes the skill emit {"action":"keep","reason":"disabled"} without calling Claude.

How

1. ffprobe the clip duration.
2. Slice the trimmed transcript text into a head window (first 1.5s) and a tail window (last 1.5s).
3. ffmpeg out a 3-frame strip from each end (start / mid / end of the 1.5s window).
4. Send both strips + both transcript snippets to claude -p (vision-enabled) with explicit inward-only rules.
5. Validate the reply: enforce 0 <= t0, t1 <= dur, t0 >= 0, t1 > t0. If the proposed trim would remove more than 2.0s of clip duration, downgrade to {"action":"drop"}. If the proposed trim would shrink the short below 15s, downgrade to {"action":"keep","reason":"would shrink below 15s"}.

Caller contract

shorts.sh reads the JSON. If action=trim, it issues a second cut-clip + rebase against the trimmed transcript, then continues to fit-vertical. If action=drop, it continues the per-span loop and increments the skipped counter. If action=keep, no-op.

Idempotency

Keyed off (clip mtime, trimmed-transcript mtime). Cached decision written next to <out>.vbmeta.