Video to shorts

Video To Shorts

Dependencies

/video-understand is a prerequisite. Run it first so shorts use the validated source transcript and canonical timeline. Before starting, verify that it is installed. If it is not, warn the user that this prerequisite is missing and stop before processing media.

Require ffmpeg/ffprobe on PATH and Python with Pillow. Check them before processing media.

Scope

This skill owns short candidate judgment, deterministic qualification, horizontal extraction, short-relative transcripts, and optional deterministic vertical delivery. It does not transcribe in project mode, change the main sequence, add captions/graphics, grade, publish, or perform continuous subject tracking.

Shorts are derivatives of final/final-video.mp4. Record the operation in the project DAG, but never add it to sequences.main.operations; the main delivery and its original-vs-final comparison remain unchanged.

Project Inputs

• work/project.json, whose main render status is verified
• final/final-video.mp4, exactly matching the project render path
• work/understand/transcript.json in source time
• work/timeline.json
• active main-operation revisions

prepare_transcript.py --timeline calls the shared projectlib.map_transcript_to_timeline. Candidate/extraction times are therefore program seconds on the actual final video. Words removed by the main edit are not available to shorts. Each retained word keeps its source_range, program_range, and clip_id.

Project transcript metadata binds the final video, source transcript, and timeline by absolute path, size, modified time, and SHA-256. Candidate validation refuses a changed binding.

Output Layout

work/shorts/
|-- transcript.json
|-- transcript_metadata.json
|-- candidates.json
|-- shorts-plan.json
|-- review/
|   |-- candidate_review.json
|   `-- approved_candidates.json
`-- short-001/
    |-- transcript.json
    |-- extraction-report.json
    `-- vertical-agent/
        |-- agent_vertical_plan.json
        |-- vertical_plan.json
        `-- review/vertical_review.json

work/cache/shorts/                 # disposable frames/sheets/intermediate media

review/06-shorts/
|-- candidates-<review-id>.html              # immutable authoritative page
|-- candidates-<review-id>-question.md
|-- assets/candidates-<review-id>/            # immutable start/middle/end frames
|-- candidates.html                           # non-authoritative latest alias
|-- candidates-summary.md
|-- short-001-horizontal.jpg
|-- short-001-vertical-preview.mp4
|-- short-001-vertical-contact-sheet.jpg
|-- short-001-vertical-preview-summary.json
|-- short-001-vertical-preview-probe.json
|-- <short-id>-vertical-review-<review-id>.html # immutable authoritative page
|-- <short-id>-vertical-review-<review-id>-question.md
|-- <short-id>-vertical-review.html             # non-authoritative latest alias
|-- <short-id>-vertical-review-assets/
|   `-- <review-id>/                             # immutable authoritative evidence
|       |-- preview.mp4                          # absent for REVIEW_REQUIRED
|       |-- contact-sheet.jpg                    # absent for REVIEW_REQUIRED
|       |-- preview-summary.json
|       `-- media-probe.json
`-- shorts-summary.md

final/shorts/
|-- short-001-horizontal.mp4
`-- short-001-vertical.mp4          # only when selected and approved

Program Transcript

Run from the repository root:

$RepoRoot = (Resolve-Path ".").Path
$SkillRoot = Join-Path $RepoRoot "skills\video-to-shorts"
$ProjectRoot = (Resolve-Path "<project-root>").Path
$Work = Join-Path $ProjectRoot "work"
$ShortsWork = Join-Path $Work "shorts"
$FinalVideo = Join-Path $ProjectRoot "final\final-video.mp4"
$ShortsReview = Join-Path $ProjectRoot "review\06-shorts"
New-Item -ItemType Directory -Force $ShortsWork,$ShortsReview | Out-Null

python "$SkillRoot\scripts\prepare_transcript.py" $FinalVideo `
  --transcript "$Work\understand\transcript.json" `
  --timeline "$Work\timeline.json" `
  --out $ShortsWork

Project timeline mode requires the canonical supplied transcript. Standalone mode may omit --timeline; it keeps input-relative time and may use the temporary video-understand/scripts/transcribe.py fallback. No transcript is auto-discovered.

Candidate Authoring

Generate long-video visual context into cache, inspect every sheet needed for a decision, and hand-author candidate JSON. Deterministic code validates editorial judgment; it does not invent candidates or scores.

python "$SkillRoot\scripts\prepare_visual_context.py" $FinalVideo `
  --out "$Work\cache\shorts\candidate-visual" --interval 15

Candidate input remains shorts-candidates.v2 for compatibility. Set selection.evidence_mode and every candidate evidence_mode to text_visual. Each candidate requires:

• stable candidate_id, title, and scene type;
• complete start_time/end_time in program time;
• transcript_excerpt exactly equal to overlapping transcript words, not whole overlapping segments;
• all six score entries and non-empty reasons;
• non-empty metadata.editorial_reason;
• optional reviewed filler_drop_spans and visual evidence fields.

Do not supply total score; the validator adds the six dimensions:

Dimension	Maximum
hook	20
completeness	20
audience_value	20
emotion_tension	15
quotability	15
pace_editability	10

Prefer a complete, independently understandable thought over a shorter fragment. Keep a dependent interviewer question, premise, claim support, reversal, or payoff when it is necessary. Boundaries must land on whole words and thoughts.

Validate the authored file:

$CandidateDir = Join-Path $ShortsWork "preview\text_visual"
python "$SkillRoot\scripts\candidates.py" `
  --out $CandidateDir `
  --candidates "$CandidateDir\candidate-input.json" `
  --transcript "$ShortsWork\transcript.json" `
  --project-root $ProjectRoot

This writes the compatibility preview files, durable work/shorts/candidates.json, and review/06-shorts/candidates.{html,summary.md}.

Review Modes

Candidate selection and delivery mode are one hash-bound gate. Delivery is exactly horizontal_only or horizontal_and_vertical.

Open the bound interactive page in human mode:

python "$SkillRoot\scripts\interaction.py" candidate-open --out $ShortsWork --review-out $ShortsReview

The command prints an authoritative review-ID-scoped page. The receipt binds that page, its candidate JSON, transcript, source media, fixed question, and every extracted frame by SHA-256. candidates.html is only a non-authoritative convenience alias to the latest page. Open the printed authoritative path with the current OS native command:

# Windows PowerShell
Start-Process (Resolve-Path -LiteralPath "<printed-authoritative-page>")
# macOS
open "<printed-authoritative-page>"
# Linux
xdg-open "<printed-authoritative-page>"
# Agent: Present the authoritative page and STOP.

If opening fails, retry the native command and report the failure if it still does not open. Do not continue after presenting it. The page starts with all candidate checkboxes and both delivery choices initially unselected. The user must select 1-5 candidates and one delivery mode. The bound page never silently selects candidates.

For approval, copy this exact page-generated structured summary unchanged into candidate-answer:

Shorts candidate review
Review: <review-id>
Candidates: text_visual/cand-001, text_visual/cand-002
Delivery: horizontal_and_vertical

For revision, copy this exact page-generated structured summary unchanged:

Shorts candidate review
Review: <review-id>
Decision: revise
Changes: <non-empty requested changes>

Record the later response without rewriting it:

python "$SkillRoot\scripts\interaction.py" candidate-answer `
  --out $ShortsWork --response-file "<verbatim-summary-file>"

When the user explicitly delegates decisions:

python "$SkillRoot\scripts\interaction.py" candidate-open `
  --out $ShortsWork --review-out $ShortsReview --decision-mode agent `
  --delegation-note "User delegated shorts selection and vertical approval."

python "$SkillRoot\scripts\interaction.py" candidate-agent-approve `
  --out $ShortsWork `
  --candidates "text_visual/cand-001" `
  --delivery horizontal_and_vertical `
  --rationale "The selected range has a self-contained hook, explanation, and payoff."

The Agent opens and inspects the same authoritative page, uses explicit candidate references, an explicit delivery mode, and a non-empty rationale. It never writes a fake human response or user_response. Human and Agent modes are mutually exclusive; human answer commands fail in Agent mode and Agent approval commands fail in human mode. Receipt verification rechecks review, page and media hashes. Regenerated source, candidate, transcript, page, question, frame, or approved-candidate artifacts invalidate approval.

Canonical Plan

python "$SkillRoot\scripts\plan.py" `
  --out $ShortsWork --project-root $ProjectRoot

work/shorts/shorts-plan.json uses integer schema_version: 1. It records:

• target: derived, timebase: program, and timeline_id;
• the verified main-render path and fingerprint;
• exact depends_on / based_on revisions from the active main sequence;
• selection mode, rationale, delivery mode, and review ID;
• stable short IDs;
• explicit program_range and mapped source_ranges;
• scoring/editorial evidence, approved filler drops, and program-time keep spans;
• work transcript/report paths and ../final/shorts/*.mp4 delivery paths.

Qualification remains deterministic: score at least 70, completeness at least 15, estimated output duration 20-90 seconds, valid exact excerpt, no greater-than-50% overlap with a higher-ranked short, and at most five outputs unless configured.

Only isolated, semantically empty hesitation words with safe word boundaries may be approved as filler drops. Rejected/unsafe spans remain in the plan with reasons. Total removal may not exceed 15% and no keep span may be shorter than 0.15 seconds.

Horizontal Extraction

python "$SkillRoot\scripts\extract_shorts.py" `
  --video $FinalVideo --out $ShortsWork `
  --project-root $ProjectRoot

Before extraction, the script rechecks the main-render fingerprint, project render status, dependency revisions, candidate receipt, approved candidate hash, and source video. Each keep range becomes its own seeked -ss/-t/-i input; ffmpeg concatenates those inputs and performs one H.264/AAC encode. It does not decode the complete main video once per keep range.

The selected final transcript word and the media cut endpoint are separate boundaries. Extraction preserves at least 0.25 seconds of release audio after selected content and targets 0.30 seconds. --no-refine-boundaries disables semantic phrase snapping only; it never disables the mandatory release handle. If the source ends before the required handle can be preserved, extraction fails instead of marking the short verified.

The short-relative transcript excludes dropped words, shifts later words onto a continuous zero-based clock, and retains input program/source evidence. Horizontal media goes directly to final/shorts; work contains only its transcript and report.

Vertical Planning

Vertical delivery starts from the approved horizontal output. Generate dense visual evidence at 1 second, or 0.5 seconds when motion requires it. Inspect every sheet and author one of:

• STATIC_CROP: one stable fixed 9:16 crop.
• SCENE_CROP: scene-bounded fixed crops, with LETTERBOX only for scenes needing complete horizontal information.
• LETTERBOX: complete sharp foreground over a blurred/darkened fill.
• REVIEW_REQUIRED: evidence cannot support deterministic fixed crops; formal render is blocked.

Presenter scenes should keep the head and torso stable and large. Slides, products, groups, and diagrams must retain essential information. Do not claim continuous tracking. Crop changes belong at scene boundaries.

Each plan segment fully covers the short timeline, has a strategy, content type, non-empty rationale, and valid in-bounds crop when applicable. Required content types are PRESENTER, WIDE_INFORMATION, PRODUCT, MULTI_SUBJECT, and OTHER.

$ShortId = "short-001"
$Horizontal = Join-Path $ProjectRoot "final\shorts\$ShortId-horizontal.mp4"
$VerticalWork = Join-Path $ShortsWork "$ShortId\vertical-agent"
$ExtractionReport = Join-Path $ShortsWork "$ShortId\extraction-report.json"

python "$SkillRoot\scripts\prepare_visual_context.py" $Horizontal `
  --out "$Work\cache\shorts\$ShortId-vertical-visual" --interval 1

python "$SkillRoot\scripts\vertical_plan.py" `
  --video $Horizontal `
  --input "$VerticalWork\agent_vertical_plan.json" `
  --out $VerticalWork `
  --source-video $FinalVideo `
  --extraction-report $ExtractionReport

vertical_plan.json preserves exact FPS as {num, den} and deterministically validates geometry, coverage, strategy duration, and warnings. The direct-render binding hashes the verified main delivery and the exact extraction report. Preview review still uses the approved horizontal short timeline, while formal rendering maps that timeline through the recorded keep_spans and reads pixels directly from the verified main delivery.

Render a smaller preview into the user-facing review directory:

python "$SkillRoot\scripts\render_vertical.py" `
  --video $Horizontal --plan "$VerticalWork\vertical_plan.json" `
  --out $VerticalWork --review-out $ShortsReview --mode preview

The command first writes flat latest convenience outputs named <short-id>-vertical-preview.mp4, <short-id>-vertical-contact-sheet.jpg, <short-id>-vertical-preview-summary.json, and <short-id>-vertical-preview-probe.json. It then prints the authoritative <short-id>-vertical-review-<review-id>.html page. The latest <short-id>-vertical-review.html alias is non-authoritative convenience only.

The authoritative page and receipt bind immutable review-ID-scoped copies under <short-id>-vertical-review-assets/<review-id>/: preview.mp4, contact-sheet.jpg, preview-summary.json, and media-probe.json. Later previews may replace the flat convenience outputs without changing an earlier receipt. REVIEW_REQUIRED stores only preview-summary.json and media-probe.json in its scoped directory because no preview video or contact sheet is rendered.

Open the printed authoritative path with the current OS native command:

# Windows PowerShell
Start-Process (Resolve-Path -LiteralPath "<printed-authoritative-page>")
# macOS
open "<printed-authoritative-page>"
# Linux
xdg-open "<printed-authoritative-page>"
# Agent: Present the authoritative page and STOP.

If opening fails, retry the native command and report the failure if it still does not open. Do not continue after presenting it. Inspect the complete preview, contact sheet, all segments and crop/fit decisions, media probe, and every warning. The receipt hashes the page, plan, source, preview, contact sheet, preview summary, probe, and fixed question.

For human review, copy the page's exact structured summary unchanged into vertical-answer. The three possible summaries are:

Shorts vertical review
Short: <short-id>
Review: <review-id>
Decision: approve

Shorts vertical review
Short: <short-id>
Review: <review-id>
Decision: revise
Changes: <non-empty requested changes>

Shorts vertical review
Short: <short-id>
Review: <review-id>
Decision: skip

python "$SkillRoot\scripts\interaction.py" vertical-answer `
  --out $VerticalWork --response-file "<verbatim-summary-file>"

Delegated mode uses the same page and records an honest Agent decision with a non-empty rationale:

python "$SkillRoot\scripts\interaction.py" vertical-agent-approve `
  --out $VerticalWork `
  --rationale "All sampled scenes preserve the subject and essential information."

REVIEW_REQUIRED offers only revise or skip; Agent approval is blocked. A newer candidate review invalidates an older vertical receipt. Final vertical rendering is allowed only after approval. Only approved, hash-matching review media and plans can reach final rendering; revision and skip never produce a final vertical file.

Formal render revalidates every approved artifact:

python "$SkillRoot\scripts\render_vertical.py" `
  --video $Horizontal --plan "$VerticalWork\vertical_plan.json" `
  --out $VerticalWork --review-out $ShortsReview --mode final

The formal renderer uses rational num/den, one H.264/yuv420p generation from the verified main delivery, libx264 -preset slow -crf 16, and encoded continuous audio. Preview rendering remains medium / CRF 20 for faster review. The renderer checks dimensions, duration, FPS, audio presence, source/report hashes, and keep-span mapping.

Project Operation

Add one derived operation to work/project.json, but not to the main sequence:

{
  "id": "shorts",
  "skill": "video-to-shorts",
  "revision": 1,
  "depends_on": ["understanding", "captions"],
  "based_on": {"understanding": 1, "captions": 1},
  "status": "verified",
  "plan": "shorts/shorts-plan.json",
  "outputs": [
    "../final/shorts/short-001-horizontal.mp4",
    "../final/shorts/short-001-vertical.mp4"
  ],
  "target": {"sequence": "main", "scope": "derivatives"},
  "effects": {
    "changes_timeline": true,
    "changes_geometry": true,
    "changes_video_pixels": true,
    "changes_audio": true,
    "adds_track": null
  },
  "check": {"status": "pass", "report": "../review/06-shorts/shorts-summary.md"}
}

Use the actual dependency list/revisions stored in the plan. Project validation checks the DAG and files; shorts remain outside the main render plan.

Compatibility

Without --project-root, candidates and shorts-plan.v2 keep their established standalone paths and input-relative time. The extractor accepts both plan shapes. Legacy short_XX/source.mp4 vertical workflows and human review commands remain valid. Only the legacy unbound text review may use the default top five candidates; the bound interactive page always requires explicit selection. Standalone fallback transcription uses the canonical video-understand transcriber and a project-local model cache.

Self Check

Run:

python "$SkillRoot\scripts\check_project_protocol.py"
python "$RepoRoot\skills\video-understand\scripts\validate.py" project `
  "$Work\project.json" $ProjectRoot

For every delivered short verify:

• candidate excerpt equals overlapping program words and the thought is complete;
• source/program ranges and dependency revisions resolve;
• horizontal duration approximately equals keep-span sum;
• short transcript starts at zero, ends within media, excludes removed words and words heard only inside the release handle, and reports tail_release_verified: true;
• horizontal output has expected FPS, H.264 video, synchronized audio, and review still;
• vertical output is 9:16, uses the exact rational FPS, keeps required content safe, has synchronized audio, matches its approved preview/plan hashes, and records direct_render: true with the verified main delivery as render_source;
• all user-facing media is under review/ or final/, and disposable material is under work/cache/;
• shorts-summary.md records outputs, decisions, warnings, and checks.

Page generation or JSON validation alone is not success. Watch the horizontal and vertical outputs and inspect their review images/contact sheets.