reel-reviewlisted
Install: claude install-skill vaibhav-barman/live-clipper
# REEL REVIEW — the final gate
> **THE SPLIT:** a script measures what is *measurable*; an agent judges what is *judgement*. Never let
> the agent guess at numbers it can measure, and never let the script rule on taste. Both halves run,
> every time.
**A second model reviews the first model's work** — that is the point. If the editor agents run on one
model, the quality-critic runs on another. Never review your own edit inline.
Adapted from the creator's gate. **Kept:** the split, the different-model rule, BLOCKER/FIX/NIT, the
two sample points per section, and how to read the numbers honestly. **Added:** the script itself
(`scripts/review_reel.py`, which the original skill used but did not ship), rebuilt for stream clips.
## Step 1 — deterministic pass
```bash
python3 scripts/review_reel.py out/<clip>.mp4 --out work/<stream>/<clip>/review \
--cut work/<stream>/<clip>/cut.json \
--captions work/<stream>/<clip>/captions.json \
--voice work/<stream>/<clip>/cut_voice16k.wav \
--overlay work/<stream>/<clip>/overlay.mov --face-band 1180:1820 \
--region game:1080:726:0:0 --region face:1008:1050:36:870
```
Every flag after the video is optional — a check whose input is missing is reported `SKIP`, never PASS.
Exit code **1 on any FAIL**, 0 otherwise. Output: `report.json`, `report.md`, `sheet_mid.jpg`,
`sheet_post.jpg`.
| check | what it catches |
|---|---|
| specs | wrong size/fps/codec/pix_fmt, audio not AAC 48 kHz stereo, duration outside 5–60 s |
| cfr | VFR outpu