Opus 5.5 edits video in Claude Code: one prompt, 15 jobs, 100/100 QA score
socialwithaayan · x · 2026-10-19
- A full guide to auto-editing video with Claude Opus 5.5 inside Claude Code: the model never watches the footage — it edits words via a Whisper transcript, not pixels.
- Test case: one raw 45-second take full of mistakes plus a single prompt; the model did 15 jobs on its own — cut mistakes, found B-roll, synced zooms to speech, burned captions, added music — and returned a 26.5s finished video saved to Google Drive, scoring 100/100 on QA.
- The guide breaks the method into 10 steps, each with the exact prompt, expected output, and verification; a final prompt runs everything at once.
- Key technique and QA gate: cut on true silence, not Whisper's word edges, or words get clipped ("Log code reads" vs "Claude Code reads" after re-transcription without a glossary). The old cut left 0.62s of silence and scored 95; the silence cut leaves at most 0.25s and scores 100 (gate: 90).
- Audio standards: -14 LUFS, audio length matched to video, ending within 0.2s of the last word; delivery spec H.264, yuv420p, tv range, faststart.
More from coding & agent
- Anthropic Built Scion, Its Own Source Control System, as Merged Changes Grew 12x to 30,000/Day — Pavan_Belagatti · 2026-10-19
- No editing software: one creator made a full AI video just by chatting with Claude — aziz4ai · 2026-10-19
- 20+ exposed panels reveal open-source AI hacking agent Cairn in real cyberattacks — Hesamation · 2026-10-19
- The Vibe Coding Curse: Understanding AI Code Defeats the Speed, Not Trusting It Blocks Shipping — MoonL88537 · 2026-10-19
- Six months of agent-memory provenance fields nobody could see: a Reddit stranger found the gap — Neither-Witness-6010 · 2026-10-19
- The AI coding curse: skipping code understanding makes you helplessly dependent on the model — MoonL88537 · 2026-10-19