stringout-for-review

I'd like you to add a skill to Wideframe called "stringout-for-review".

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

---
name: stringout-for-review
description: Transcribe interview or talking-head footage and lay every usable spoken segment end to end on one timeline with a marker at each soundbite, delivered as a validated Premiere Pro project plus a review MP4. Use when someone says "stringout", "selects reel", "give me everything they said about …", "a timeline I can review in Premiere", or wants to scrub the good moments before deciding on a cut.
argument-hint: "[path-to-footage] [optional: topics or people to mark]"
---

# Stringout for Review

A review stringout places usable spoken takes in source order, with markers for navigation. This skill uses the silence and minimum-length thresholds below to build a Premiere project and a review MP4; keep the full source available when those thresholds could exclude important short replies.

## Scope

Assembly and marking, not editing. Takes are kept whole in source order. Filler removal is optional and off by default for a review stringout because reviewers want to hear the real delivery.

## Step 0: Resolve the input and set defaults

Accept a folder, files, or a workspace. If the user named topics, people or questions to find, keep that list; it drives the markers in Step 3.

Defaults unless told otherwise:

- Include every speaker.
- Verbatim (no filler removal).
- Split a take when silence exceeds 1.5 seconds; drop runs shorter than 2 seconds.
- One-second gap between takes so cuts are visible when scrubbing.
- Sequence settings match the majority of the footage.

Ask nothing yet.

## Step 1: Transcribe everything

1. `caltools media-info-batch` for fps, resolution, duration and audio presence per clip.
2. `caltools media-transcript-get-batch` on every clip with audio, word-level timestamps.
3. `caltools job-poll` until nothing is pending.

If speaker labels are available, keep them; if the footage has more than one speaker and labels are missing, say so and continue rather than guessing names.

Evaluation: every clip with audio has a transcript. Clips without speech are listed so the user knows they were seen and left out.

## Step 2: Segment into takes

Split each transcript into spoken runs using the silence threshold. Each run becomes one take with a source clip, in and out timecodes, and its text. Order takes by source clip and then by time, unless the user asked for a topic order.

Evaluation: the sum of take durations plus gaps equals the planned sequence length, and no take is shorter than the minimum.

## Step 3: Decide the markers

Two marker sets, both applied:

1. One marker per take, named with the clip name and the first six words of the take. Color: default.
2. If the user gave topics or people: a marker at every mention, literal or by meaning, named `<topic>: <first words>`, in a second color. Merge hits within 5 seconds into one marker.

Evaluation: marker count equals take count plus post-merge topic hits, and every marker maps to a timecode inside a take.

## Step 4: Build the sequence

1. `caltools sequence-create` at the majority frame rate and resolution.
2. `caltools sequence-media-set` so settings match the footage.
3. `caltools sequence-clips-add` each take in order with the confirmed gap; `sequence-clips-update` to set exact in and out points.
4. `caltools sequence-markers-add` for both marker sets.

Evaluation: render frames at three points with `sequence-render-frames`; each shows a speaker, none show black.

## Step 5: Export, validate, render

1. `caltools sequence-export-premiere` to `<workspace>/renders/<name>_stringout_v1.prproj`.
2. `caltools prproj validate`. Zero dangling references and orphans, or fix and re-run.
3. `caltools prproj inspect`: clip count equals take count, marker count equals Step 3.
4. `caltools sequence-render-final` to `<workspace>/renders/<name>_stringout_v1.mp4`. Confirm it plays with audio via `media-info-batch`.

## Step 6: Report and hand off

Give the user: the MP4, the .prproj, a table of takes (number, clip, timecode in the stringout, first words, and any topic tags), and the list of clips with no speech. Then ask one question: "Want me to cut this down to a first assembly, or add captions?" Name `rough-cut-to-premiere` and `caption-this` as the next steps.

## Done when

- Every spoken take from every clip appears exactly once, in order, with a marker.
- Topic markers exist for every requested topic, with per-topic counts reported.
- The `.prproj` passed `prproj validate` and its clip and marker counts reconcile.
- A review MP4 exists, plays, and has audio.