platform-cutter

---
name: platform-cutter
description: Cuts one recording into native TikTok, Reels, Shorts and LinkedIn versions, each with its own encode, caption geometry and copy. Use after a shoot when the footage goes out everywhere.
---

# One shoot, cut correctly for TikTok, Reels, Shorts and LinkedIn

Four native uploads from one recording, each encoded and written for its app.

## Claude does

1. **Collect the inputs.** Source path, transcript with timecodes, the platforms, the goal for each (completions, saves, comments, shares), the account posting it. No timecodes, no cut: say so and stop.
2. **Atomise.** 5 to 10 self-contained atoms, each making sense to someone who saw nothing else.
3. **Score every atom** on the claude-shorts rubric: hook 0.30, coherence 0.25, emotion 0.20, value density 0.15, payoff 0.10. Under 60 is not cut.
4. **Assign atoms to platforms.** Vertical cuts run 30 to 60 seconds; the claude-shorts examples land at 39 to 43. LinkedIn gets a native upload that makes its point before the fold.
5. **Rewrite copy per platform, never reuse it.** TikTok: up to 2,200 characters, 3 to 5 hashtags of mixed reach, hashtags last. Reels and Shorts get their own openers. LinkedIn gets prose, not hashtag stacking.
6. **Fix caption geometry.** 1080x1920 everywhere; burned-in text stays inside a centred 900x1400 box, since app furniture differs per platform and moves. Flag it as a spec to confirm in each preview.
7. **Output the clip table:** clip_id, platform, in and out timecodes, target length, spoken hook, on-screen text.

## Then Codex does

The mechanical half: frame-accurate cutting, one encode per platform, a manifest. Bulk ffmpeg over a decided table.

```
codex exec --sandbox danger-full-access --skip-git-repo-check -C "<working folder>" - < CODEX.md
```

Fill in first: {SOURCE_VIDEO}, {CLIP_TABLE}, {OUTPUT_DIR}, {ACCOUNT_NAME}.

## Claude checks

1. Play the first and last second of every clip. Roughly 70% of auto-cut boundaries clip a word; a half-word open or ending is a reject.
2. Read specs from ffprobe, not Codex's prose. Shorts H.264 High 4.2, 12 Mbps, AAC 192k. TikTok CRF 18, AAC 128k. Reels 4.5 Mbps maximum, AAC 128k. All 1080x1920.
3. Watch one clip per platform muted. If it does not stand alone, it fails.
4. No burned-in text outside the centred 900x1400 box.
5. On failure, name the clip_ids and rerun Codex. Never hand-patch the manifest.

## Rules

- One export cannot serve four platforms: re-encode and rewrite for each.
- Never stream-copy a cut: it snaps to the nearest keyframe and eats the first word.
- Voice: no em dashes, no "leverage", "delve" or "hey guys". Specifics beat adjectives ("3 takes", not "a few").
- Write for muted viewing: 80% say captions make them likelier to finish a video.
- Public information only.
- Never invent a fact, a number or a quote.
- Anything sent in someone's name says whose name it is.

## If Codex is not installed

Claude does the cuts and encodes itself, one clip at a time, and says so.

## Built from

- claude-shorts, https://github.com/AgriciDaniel/claude-shorts, 204 stars (api.github.com, 5 Sep 2026): rubric, encode table.
- social-media-skills/skills, https://github.com/social-media-skills/skills, 66 stars: atomising, cleanup figure.
- claude-youtube, https://github.com/AgriciDaniel/claude-youtube, 350 stars: /youtube repurpose.
- tiktok-skills, https://github.com/sergebulaev/tiktok-skills, 22 stars: caption spec, voice rules.
- 3Play Media on the Verizon and Publicis captions study, https://www.3playmedia.com/blog/verizon-media-and-publicis-media-find-viewers-want-captions/: 5,616 adults, the 80% figure.

# One shoot, four platform cuts - Codex task

## You are given

- **{SOURCE_VIDEO}** - absolute path to the master recording. Read-only.
- **{CLIP_TABLE}** - one row per clip: clip_id, platform (tiktok | reels | shorts | linkedin), in_timecode, out_timecode, target_length_seconds, on_screen_hook. The timecodes are already decided. Do not move them.
- **{OUTPUT_DIR}** - folder for the exports. Create it and a subfolder per platform.
- **{ACCOUNT_NAME}** - the account these will be posted from. Used in every filename and in the manifest.

## Produce

For every row, one MP4 at `{OUTPUT_DIR}/<platform>/{ACCOUNT_NAME}_<clip_id>_<platform>.mp4`:

- All platforms: 1080x1920, H.264 video, AAC audio, audio normalised to -14 LUFS integrated with true peak at -1 dBTP.
- **shorts**: H.264 High profile, level 4.2, 12 Mbps video, AAC 192 kbps.
- **tiktok**: CRF 18, AAC 128 kbps.
- **reels**: 4.5 Mbps video maximum, AAC 128 kbps.
- **linkedin**: use the reels settings and record in the manifest that no separate published spec was supplied.

Also produce, per clip, a 1080x1920 PNG of the first frame at `{OUTPUT_DIR}/<platform>/<clip_id>_frame1.png`, and one `manifest.json` plus `MANIFEST.md` at the root of {OUTPUT_DIR}.

## Rules

- Re-encode every cut. Never `-c copy`: a copy cut snaps to the nearest keyframe and swallows the first word.
- Seek accurately (input seek followed by an output-side trim, or `-ss` after `-i`), then verify the written duration against out minus in. Tolerance is 0.04 seconds.
- If the source is not 9:16, scale to fit 1080x1920 and pad with black. Do not crop faces out of frame, and do not guess a crop window.
- Never adjust an in or out point yourself. If a clip fails, leave it out and report the clip_id.
- Do not write captions, hashtags, hooks or any copy. Text is decided elsewhere.
- Do not burn in subtitles unless the clip table supplies a subtitle file path.
- Every value you report is read back from `ffprobe` on the written file. Never report the setting you asked for.
- Do not upload, post or publish anything. Do not modify or delete {SOURCE_VIDEO}.
- Overwrite an existing export only if the clip_id matches; never clear the folder.

## Return

Print one table, one row per written file, with these columns read from ffprobe: clip_id, platform, output path, duration, width x height, video codec with profile and level, video bitrate, audio codec with bitrate, integrated loudness.

Then print a `FAILED` block listing any clip_id that did not produce a file, with the exact ffmpeg error line for each.

Then print `TOTAL: <n> written, <n> failed`.

Print nothing else. No summary, no advice, no next steps.