---
name: social-video-captioning
description: Use when Jared sends short social videos and asks for captions, caption trials, post copy, hooks, variants, or asks whether a recorded explanation is correct. Covers direct video inspection, transcript grounding, punchy caption drafting, and concise explanation critique.
---

# social-video-captioning

Use this skill when Jared sends one or more short videos, social posts, X/Twitter links, or creator content and asks for captions, talking points, post copy, hooks, options for testing, or whether his recorded explanation is correct and clear.

## Operating rule

Do not ask Jared to describe attached videos. Inspect them directly.

The job is not to summarise the video. The job is to create usable captions that match the video's argument and are ready to post or test.

## Workflow

1. **Inspect the source first.** For uploaded video files, use `ffprobe` to confirm duration, dimensions, and whether audio exists. For X/Twitter links, extract the post text and any attached media before drafting.
2. **Create visual grounding.** Make a contact sheet with `ffmpeg` at roughly one frame per second for videos. Identify the subject, setting, on-screen text, gestures, mood, and visual hook.
3. **Transcribe when audio exists.** Use `hyperframes transcribe` so the caption or talking points reflect what is actually said, not what the frames imply.
4. **Fact-check if asked whether it is true.** When Jared asks to verify a recorded explanation, extract claims from both the transcript and on-screen text, then verify against primary sources before giving a publish-readiness verdict. Use `references/video-explanation-fact-checking.md`.
5. **Turn the source into usable output.** If Jared asks for talking points, give the strategic angles, plain-language soundbites, audience-specific implications, and watch-outs. Do not just summarise the post.
6. **Draft distinct trials.** If Jared says he is testing captions, produce genuinely different angles, not tiny rewrites.
7. **Keep the output lean.** Give the captions, talking points, or verdict. Only add a short strategic note if one version is clearly stronger.

## Jared's default social-caption style

- Punchy.
- Active voice.
- Short lines.
- Opinionated.
- Human, not corporate.
- No em dashes.
- No long setup.
- No explaining the process.

## Default output shape

For two videos, use:

```markdown
## Video 1: [short working title]

### Caption trial one
[caption]

### Caption trial two
[caption]

## Video 2: [short working title]

### Caption trial one
[caption]

### Caption trial two
[caption]
```

If one video is stronger, say it plainly in one line.

## Useful commands

```bash
ffprobe -v error -show_entries format=duration,size:stream=index,codec_type,codec_name,width,height,r_frame_rate -of json "$video"
ffmpeg -y -i "$video" -vf "fps=1,scale=232:-1,tile=4x4" -frames:v 1 /tmp/video_sheet.jpg
hyperframes transcribe --engine auto --model base.en --language en --json -d /tmp "$video"
```

## Pitfalls

- Do not rely only on the contact sheet when the video has speech.
- When transcribing more than one video, run `hyperframes transcribe` separately for each file into distinct output directories. A multi-file call can return a single `transcript.json`, which makes it unclear which clip the transcript belongs to.
- When Downloads contains duplicate exported files like `Video.mp4` and `Video (1).mp4`, inspect sizes and durations first, then use the clean non-duplicate filename unless the user clearly names the copy.
- Do not over-polish into brand manifesto copy unless Jared asks for that.
- Do not give ten variants when he asked for two trials.
- Do not make the caption a literal transcript. Use the transcript to find the argument, then write the post.
- When Jared pushes back on caption framing as too generic, stop expanding the concept and simplify the mechanics. For AI/folder/markdown explainer posts, the core mechanic is: AI only uses what enters its context window; folders organise the source material; markdown is plain text with headings the AI can parse; retrieval pulls the relevant sections into the answer. Avoid "business brain" language unless he asks for a strategic frame.
- When Jared asks whether HyperFrames is the best local editing software for creator reels, do not oversell it as a universal editor. Position it as a repeatable production engine, not the easiest manual timeline editor. CapCut Desktop, DaVinci Resolve, Final Cut Pro, or Premiere are better for hands-on cuts and finishing polish; HyperFrames is better for reusable 9:16 systems with captions, PiP, screenshots, proof cards, and batch variation.
- For AI creator reels in the style of a YouTube Short with talking head, proof screens, and cinematic b-roll, recommend the practical stack: HeyGen for face/avatar source, Higgsfield for pattern-interrupt b-roll, HyperFrames for composition/render, and CapCut or DaVinci only as a final polish bench.

## References

- `references/social-video-caption-trials.md` for the detailed inspection and drafting pattern.
- `references/video-explanation-review.md` for checking whether Jared explained an idea correctly from a recorded video.
- `references/video-explanation-fact-checking.md` for verifying factual claims in a recorded social video and giving a publish-readiness verdict.
- `references/x-post-extraction.md` for extracting X/Twitter posts, attached videos, and drafting talking points when normal scraping fails.