social-clips

I'd like you to add a skill to Wideframe called "social-clips".

Save the SKILL.md below exactly as written, then confirm I can run it by saying /social-clips.

---
name: social-clips
description: Find the best moments in longer footage or a finished cut and turn them into short vertical clips with captions, each delivered as its own MP4 plus one validated Premiere Pro project holding every clip sequence. Use when someone says "Instagram clips", "Reels", "TikToks", "Shorts", "cuts for LinkedIn", "viral moments", "the best 30 seconds", or wants a panel, podcast, interview, event or sermon repurposed for social.
argument-hint: "[path-to-footage-or-sequence] [optional: platform, count, length]"
---

# Social Clips

This skill finds self-contained moments, reframes them vertical with the subject kept in frame, captions them so they work on mute, and delivers one MP4 per clip and one Premiere project with a sequence per clip. It makes initial editorial choices for the user to review and asks one question at the end.

## Scope

Short clips from existing footage. Not a re-edit of the long piece, not thumbnails, not posting. Music only if supplied.

## Step 0: Resolve the input and set defaults

Accept a folder of footage, a single long recording, a Wideframe sequence, or a `.prproj` plus sequence name. Read any brief (platform, audience, count, length).

Defaults unless told otherwise:

- Platform: Instagram Reels, 9:16 at 1080×1920, 30 fps or the source rate if it is 24 or 25.
- Count: 5 clips. Length: 20 to 45 seconds each, hard cap 60.
- Every clip stands alone: a complete thought with a start and an end, no mid-sentence openings.
- Captions burned in, using the `caption-this` standard (2 lines, 32 characters, 17 characters per second).
- Subject-aware reframe: keep the speaker's face in the upper two thirds with room for captions below.

Ask nothing yet.

## Step 1: Understand the footage

1. `caltools media-info-batch` for fps, resolution, rotation, duration and audio.
2. `caltools media-transcript-get-batch` with word-level timestamps on everything with speech.
3. `caltools media-analyze-content-batch` for scene, subject and energy per range (laughter, applause, movement, a second speaker joining).
4. `caltools job-poll` until nothing is pending.

Evaluation: full transcript coverage and a content summary per clip; unreadable files listed.

## Step 2: Pick the moments

Score candidate ranges on: a complete self-contained idea, a strong first sentence (the hook), emotion or energy from Step 1, and no dependence on context the viewer will not have. For a brief like "insightful thoughts about X", match by meaning, not keywords. Prefer variety across the set: different topics, and if there are several speakers, more than one of them.

Write the pick list: clip number, source, in and out, the hook sentence, why it was chosen. Trim each range to start on the first word of the hook and end on the last word of the thought, with a beat of air on each side.

Evaluation: the requested count of picks, each within the length window, none overlapping, each starting on a sentence boundary.

## Step 3: Build one clip, check it, then batch

For the first pick only:

1. `caltools sequence-create` at 1080×1920 and the chosen frame rate; `sequence-media-set` to match.
2. `caltools sequence-clips-add` the range; `sequence-clips-update` for the exact in and out.
3. Reframe: `caltools media-analyze-content-batch` to track the subject across the range, then apply a crop path through the clip transform that holds the face in the safe area. Where the subject leaves frame, hold the last good position rather than snapping.
4. If the source cuts between angles inside the range, keep the cuts and reframe each segment on its own subject.
5. Captions: build cues per the `caption-this` standard and add them with `sequence-text-overlay-add`, positioned in the lower third above the platform UI zone.
6. `caltools sequence-transitions` only at the head and tail (a short fade), never inside a clip.
7. `sequence-render-frames` at the start, middle and end: face in the safe area, captions legible and inside the frame, no black bars, no stretch.

If the first clip passes, repeat for the rest with the same settings. If it fails, fix the rule before batching.

Evaluation: every clip sequence renders three clean check frames.

## Step 4: Export, validate, render

1. `caltools sequence-export-premiere` with every clip sequence in one project at `<workspace>/renders/<name>_social_v1.prproj`, one bin per clip.
2. `caltools prproj validate`; fix and re-run on any error. `prproj inspect`: sequence count equals clip count.
3. `caltools sequence-render-final` each clip to `<workspace>/renders/social/<nn>_<hook-slug>.mp4`, H.264, 1080×1920. Confirm each with `media-info-batch`: exact dimensions, audio present, duration within the window.

Evaluation: one MP4 per pick, all at the target dimensions, and the project validates.

## Step 5: Report and ask the one question

Give a table: clip number, hook sentence, length, MP4 path, and the source timecode. Give the .prproj path. State the defaults you applied (platform, count, length, caption standard). Then ask one question: "Want more clips, a different platform size, or changes to any of these?"

Name the follow-ups: `reframe` for 1:1 and 4:5 versions of the finished clips, and `caption-this` if they want an SRT instead of burned-in text. Add either recipe from the public skills library if it is not already saved.

## Done when

- The requested number of clips exists, each a complete thought within the length window, with no overlaps.
- Every MP4 is exactly the target dimensions, plays, and has audio.
- The subject stays in the safe area and captions are legible in every check frame.
- The `.prproj` passed `prproj validate` with one sequence per clip.
- The user was asked one question, at the end.