video-add-captions-x

原始内容

---
name: video-add-captions
description: >
  Add word-timed captions to an Open Recut program. Use this skill to map the
  canonical transcript through timeline.json, review a maintained style on
  source-backed pixels, render a local transparent HyperFrames PNG sequence,
  and register it as an overlay contribution for the shared delivery render.
---

# Video Add Captions

## Dependencies

`/video-understand` is a prerequisite. Run it first so captions use the validated
word-level 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, Python with `Pillow`, and Node.js >= 22 (for the `.mjs` scripts and `npx hyperframes`, fetched on demand). `hyperframes render`/`snapshot` drives a headless Chrome — it manages its own `chrome-headless-shell`, and falls back to a system Chrome (set `CHROME` to override) when the cached one is unusable. Check these before processing media.

## Scope

This skill owns caption grouping, style selection, review, and the transparent
caption track. It does not transcribe, cut, retime, grade, reframe, or choose the
delivery audio policy.

Use `video-understand` first. If a cut exists, caption the active program
timeline; do not treat source transcript seconds as program seconds.

## Protocol Inputs

Required:

- `work/project.json`
- `work/understand/transcript.json`, with word-level source timestamps
- `work/understand/media.json`
- `work/timeline.json`
- the source video named by `work/project.json`

The timeline is the only source-to-program mapping. All ranges are half-open
`[start_s, end_s)`. `scripts/build_captions.py` uses the shared
`projectlib.map_transcript_to_timeline`; words in dropped source ranges disappear,
retained words move to program time, and cues always break at clip boundaries.

## Durable Outputs

```text
work/captions/
|-- captions-plan.json
`-- caption-interaction.json
work/cache/captions/
|-- preview-project/
|-- preview-snapshots/
|-- overlay-project/
`-- overlay-frames/frame_000001.png ...
review/05-captions/
|-- captions.srt
|-- captions-style-review-<UUID>.html
|-- captions-style-review.html
|-- captions-review.html
|-- preview-early.png
|-- preview-middle.png
|-- preview-late.png
|-- preview-no-caption.png
|-- captions-evidence.json
`-- captions-summary.md
```

Generated HTML, copied runtime files, extracted source frames, and overlay frames
are cache. The plan, SRT, decision receipt, and review evidence are durable.

## Caption Plan

`work/captions/captions-plan.json` is the canonical caption plan, targets `overlay`, and uses
program time. It records:

- `timeline_id`, `program_duration_s`, and the source transcript path;
- cue grouping settings;
- approved style, decision mode, rationale, gallery choice, and complete resolved style;
- source-backed review evidence and approval actor;
- cues with compatibility fields `start`, `end`, `text`, `lines`, and `words`;
- `program_range` and `source_ranges` on every cue;
- program/source ranges and `clip_id` on every mapped word;
- a local renderer recipe with exact rational FPS and hashed runtime assets.

The generator also accepts the old top-level cue array. That is a standalone
compatibility adapter only; it cannot express timeline provenance or regenerate an
approved protocol contribution.

Standard is the default presentation mode. Standard plans do not need a
`presentation` field and must not be migrated merely to add one. Expressive is an
explicit presentation mode, not a preset: request it with
`--presentation-mode expressive`. The build step then adds stable cue IDs, default
word `semantic_role` values, and a draft `presentation` planning shell; it does not
guess layout variants.

Expressive supports only `bottom-standard` and `center-emphasis`. Read
`reference/caption-rules.md` before filling the shell. The
existing renderer consumes only completed `presentation.layout_beats`; it does not
infer positions or change layout inside a beat.

## Styles

Maintained presets are `clean`, `minimal`, `social-bold`, `pill`, `boxed`,
`stroked`, and `shorts`. Gallery combination IDs add maintained themes, for
example `pill-yellow`, `stroked-blue`, and `shorts-green`. Karaoke is an option,
not a separate preset.

Default delegated choice: `clean`, Karaoke off. Choose a more expressive preset
only when the destination or footage justifies it. Store job-specific changes in
an overrides JSON; do not edit the generator for ordinary style feedback.

Expressive supports both `semantic-only` (Karaoke off) and
`semantic-plus-karaoke` (Karaoke on). Semantic emphasis persists for the full cue
in both configurations. In coexistence mode, semantic emphasis and the active
Karaoke word use `style.wordHighlight.activeColor` as their foreground color in
every highlight mode; a background highlight remains an additional effect. The
effective scale is `max(semantic scale, karaoke active scale)`, never their
product. Expressive never renders an underline, including for the compatible
`contrast` semantic role.

The existing review contract continues to use semantic-only as the primary
Expressive evidence and binds semantic-plus-karaoke separately. The compatibility
payload field remains named `experimental_comparison`; do not rename it without an
approval-protocol migration.

Read:

- `reference/caption-rules.md` before changing grouping limits;
- `reference/caption-style-themes.md` for style vocabulary;
- `reference/caption-feedback-mapping.md` for override mapping.

## Decision Modes

Two honest modes share the same hash-bound state machine:

- `human`: record the user's exact gallery response with `select`, show all real
  preview evidence, and record the exact confirmation response with `confirm`.
- `agent`: use only when the user explicitly delegates caption decisions. Start
  with a non-empty delegation note, then use `agent-select` and `agent-confirm`
  with non-empty rationales.

Human commands fail in agent mode and agent commands fail in human mode. Both
modes require four existing source-backed images before approval. A change to the
source, plan, timeline, style, override, project metadata, review page, or evidence
invalidates approval; rebuild the affected artifacts and review them again.

## Canonical Workflow

Run from the repository root. Set project paths:

```powershell
$RepoRoot = (Resolve-Path ".").Path
$SkillRoot = Join-Path $RepoRoot "skills\video-add-captions"
$ProjectRoot = (Resolve-Path "<project-root>").Path
$Work = Join-Path $ProjectRoot "work"
$SourceVideo = "<absolute path from work/project.json>"
$Plan = Join-Path $Work "captions\captions-plan.json"
$Receipt = Join-Path $Work "captions\caption-interaction.json"
$Review = Join-Path $ProjectRoot "review\05-captions"
$Cache = Join-Path $Work "cache\captions"
$PreviewProject = Join-Path $Cache "preview-project"
$ComparisonProject = Join-Path $Cache "preview-project-expressive-karaoke"
$OverlayProject = Join-Path $Cache "overlay-project"
$OverlayFrames = Join-Path $Cache "overlay-frames"
New-Item -ItemType Directory -Force -Path (Split-Path $Plan),$Review,$Cache | Out-Null
```

Before building the caption plan, ask the user to choose the presentation mode
unless the current request already explicitly selects one. Render the question in
the user's current conversation language while keeping the `Standard` and
`Expressive` identifiers untranslated. Keep the English source prompt below in
this file, present the localized choice in the current session, then
**Present + STOP**:

```text
Choose the caption presentation mode:

1. Standard: Uses stable conventional placement and suits most content.
2. Expressive: Switches between bottom and center placement based on meaning and emphasizes keywords and numbers.

Reply with Standard, Expressive, or "Use the default (Standard)."
```

Do not run `build_captions.py` until the user responds. If the user explicitly
delegates the choice in any wording, select Standard. State that Standard was
selected as the compatibility-preserving default before building the plan.

Build program-time cues and the review SRT:

```powershell
python "$SkillRoot\scripts\build_captions.py" `
  "$Work\understand\transcript.json" `
  $Plan `
  "$Review\captions.srt" `
  --timeline "$Work\timeline.json" `
  --source-transcript "understand/transcript.json" `
  --max-chars 42 --max-lines 2 --max-dur 6 --gap 0.6
```

After the user selects Standard, the command above produces the unchanged Standard
plan. After the user selects Expressive, run the same command with:

```powershell
  --presentation-mode expressive
```

Before starting style selection or generating a preview, the Agent must plan the
entire Expressive program in one pass:

1. Read the transcript, available understanding artifacts, canonical timeline,
   generated caption cues, and any real source-frame evidence needed to justify
   layout changes.
2. Fill all `presentation.layout_beats`, annotate exceptional word
   `semantic_role` values, write one rationale per beat, and write the overall
   `presentation.planner.rationale`.
3. Set `presentation.planning_status` to `complete` only after every cue is covered
   exactly once.
4. Validate the completed plan before preview generation:

```powershell
python "$SkillRoot\scripts\build_captions.py" --validate-plan $Plan
```

Do not ask the user to choose a position cue by cue. The user reviews the existing
HTML evidence and corrects only a small number of anomalous beats. Starting the
interaction after planning keeps the existing caption-plan hash binding authoritative.

Start one decision mode. For human mode, publish the maintained offline gallery
into the project review directory without asking the user to locate a file:

```powershell
$StartOutput = node "$SkillRoot\scripts\caption_interaction.mjs" start `
  --state $Receipt --source $SourceVideo --captions $Plan `
  --review-dir $Review --no-open true
$StartOutput | Write-Host
```

`start` prints the authoritative `captions-style-review-<UUID>.html` project page
path and the fixed question. `captions-style-review.html` is a non-authoritative
latest convenience alias only. The Agent must take the UUID path printed after
`Caption style review:`, assign that exact path to `$StyleReviewPage`, and open it
with the native command for the host OS:

Windows PowerShell:

```powershell
$StyleReviewPage = "<authoritative UUID page path printed after Caption style review:>"
Start-Process -FilePath (Resolve-Path $StyleReviewPage)
```

macOS:

```bash
open "$StyleReviewPage"
```

Linux:

```bash
xdg-open "$StyleReviewPage"
```

If opening fails, diagnose the command or path and retry it. Do not ask the user
to find the page. Present the opened gallery and the fixed question exactly as
printed, then **Present + STOP**. Do not continue until the human copies the
structured summary from the page. It has this form:

```text
Caption style review
Review: <UUID from the opened page>
Decision: select
Choice: pill-yellow
```

Pass the user's exact summary unchanged to `select --response`. In PowerShell, a
single-quoted here-string preserves the lines safely:

```powershell
$StyleResponse = @'
Caption style review
Review: <UUID from the opened page>
Decision: select
Choice: pill-yellow
'@
node "$SkillRoot\scripts\caption_interaction.mjs" select `
  --state $Receipt --response $StyleResponse
```

The block above shows the structure; use the user's returned block, including
their review UUID and choice. Never infer, normalize, or simulate a human
response.

For explicit delegation, create a separate Agent-mode receipt. Keep the bound
gallery, suppress the script's Windows-only auto-open, open the authoritative UUID
page with the matching native command above, and inspect it before choosing:

```powershell
$StartOutput = node "$SkillRoot\scripts\caption_interaction.mjs" start `
  --state $Receipt --source $SourceVideo --captions $Plan `
  --review-dir $Review `
  --decision-mode agent `
  --delegation-note "User delegated caption style and preview approval." `
  --no-open true
$StartOutput | Write-Host

node "$SkillRoot\scripts\caption_interaction.mjs" agent-select `
  --state $Receipt --choice clean `
  --rationale "Conservative readable treatment preserves the talking-head frame."
```

Use `agent-select` only after inspecting the bound gallery and record the real
rationale. Do not manufacture a human summary or use `select` in Agent mode.

Generate a transparent preview composition. The project contains only local GSAP
and font files and preserves rational FPS in `project-meta.json`:

```powershell
node "$SkillRoot\scripts\generate_caption_project.mjs" `
  --video $SourceVideo --captions $Plan --out $PreviewProject `
  --interaction-state $Receipt --project-root $ProjectRoot --mode preview

npx.cmd hyperframes check $PreviewProject --at 1 --timeout 10000 --no-contrast
```

For the primary Expressive review, append `--karaoke false`. Generate the supported
coexistence comparison from the same source, plan, interaction selection, and timing by changing
only the output directory and appending `--karaoke true`:

```powershell
node "$SkillRoot\scripts\generate_caption_project.mjs" `
  --video $SourceVideo --captions $Plan --out $ComparisonProject `
  --interaction-state $Receipt --project-root $ProjectRoot `
  --karaoke true --mode preview
```

For Standard, build mapped early/middle/late/no-caption evidence. This command captures
transparent HyperFrames snapshots, maps each program time through
`timeline.json`, extracts the matching original source frame, and composites the
two with Pillow:

```powershell
python "$SkillRoot\scripts\build_caption_review.py" `
  --source $SourceVideo --timeline "$Work\timeline.json" --plan $Plan `
  --project $PreviewProject --cache "$Cache\review-cache" --out $Review `
  --interaction-state $Receipt
```

For Expressive, the same builder emits one primary midpoint sample per layout beat,
one no-caption sample, and a separately bound semantic-only/Karaoke comparison:

```powershell
python "$SkillRoot\scripts\build_caption_review.py" `
  --source $SourceVideo --timeline "$Work\timeline.json" --plan $Plan `
  --project $PreviewProject --comparison-project $ComparisonProject `
  --cache "$Cache\review-cache" --out $Review --interaction-state $Receipt
```

The builder writes `captions-review.html` as the one authoritative page. Standard contains
exactly the four maintained evidence labels. Expressive contains the dynamic layout
beat evidence, filter controls, image enlargement, no-caption evidence, and the
separately bound coexistence comparison outside the primary approval count. The
Agent must open that page with the native command for the host OS; for example,
set `$EvidenceReviewPage = "$Review\captions-review.html"` and run one of:

```powershell
Start-Process -FilePath (Resolve-Path $EvidenceReviewPage)
```

```bash
open "$EvidenceReviewPage"
```

```bash
xdg-open "$EvidenceReviewPage"
```

If opening fails, diagnose and retry. Inspect every primary image for readability,
safe-area placement, clipping, word wrapping, stable beat placement, and unwanted
pixels in `preview-no-caption.png`. Standard binds the maintained four images:

```powershell
$Evidence = @(
  "$Review\preview-early.png",
  "$Review\preview-middle.png",
  "$Review\preview-late.png",
  "$Review\preview-no-caption.png"
) -join ","

node "$SkillRoot\scripts\caption_interaction.mjs" preview-ready `
  --state $Receipt --project-meta "$PreviewProject\project-meta.json" `
  --evidence $Evidence --review-page "$Review\captions-review.html" `
  --timeline "$Work\timeline.json"
```

Expressive binds the dynamic primary count and the separate coexistence pair from
`captions-evidence.json`:

```powershell
$EvidenceDocument = Get-Content "$Review\captions-evidence.json" -Raw | ConvertFrom-Json
$Evidence = @($EvidenceDocument.samples | ForEach-Object {
  Join-Path $Review $_.preview
}) -join ","
$ComparisonEvidence = @($EvidenceDocument.experimental_comparison.samples | ForEach-Object {
  Join-Path $Review $_.preview
}) -join ","

node "$SkillRoot\scripts\caption_interaction.mjs" preview-ready `
  --state $Receipt --project-meta "$PreviewProject\project-meta.json" `
  --evidence $Evidence --comparison-evidence $ComparisonEvidence `
  --review-page "$Review\captions-review.html" --timeline "$Work\timeline.json"
```

In human mode, present the opened page and its images, then **Present + STOP**.
The human copies one of these structured summaries from the page:

```text
Caption preview review
Review: <UUID from the opened page>
Decision: approve
Evidence: early, middle, late, no-caption
```

Expressive approval uses the dynamic evidence token and requires the explicit
Karaoke choice made after inspecting the bound comparison evidence:

```text
Caption preview review
Review: <UUID from the opened page>
Decision: approve
Evidence: expressive-layout-beats
Karaoke: on|off
```

```text
Caption preview review
Review: <UUID from the opened page>
Decision: revise
Changes: Raise captions 20 pixels.
```

Pass the returned summary unchanged to `confirm --response` or
`adjust --response`. Quote the multiline response safely in PowerShell:

```powershell
$PreviewResponse = @'
Caption preview review
Review: <UUID from the opened page>
Decision: approve
Evidence: early, middle, late, no-caption
'@
node "$SkillRoot\scripts\caption_interaction.mjs" confirm `
  --state $Receipt --response $PreviewResponse
```

For a `revise` summary, preserve it in `$PreviewResponse` the same way and run:

```powershell
node "$SkillRoot\scripts\caption_interaction.mjs" adjust `
  --state $Receipt --response $PreviewResponse
```

Apply the requested change, then regenerate the preview project, review page, and
all evidence before presenting the gate again. Never convert human feedback into
an Agent decision.

In delegated mode, the Agent must inspect the same `captions-review.html` page and
all primary evidence plus any coexistence comparison before recording its own rationale.
For Standard, keep the existing command unchanged:

```powershell
node "$SkillRoot\scripts\caption_interaction.mjs" agent-confirm `
  --state $Receipt `
  --rationale "All bound primary frames are readable and collision-free; coexistence comparison evidence was inspected separately when present."
```

For Expressive, pass the explicit Karaoke choice made from the comparison:

```powershell
node "$SkillRoot\scripts\caption_interaction.mjs" agent-confirm `
  --state $Receipt --karaoke off `
  --rationale "The semantic-only comparison is clearer for this footage."
```

Generate the approved overlay project. This formal render is the only step that
finalizes `style`, `review`, and hashed runtime assets in the canonical plan.
Expressive overlay generation uses the preview-approved Karaoke choice; an
explicit `--karaoke` value must match that approval:

```powershell
node "$SkillRoot\scripts\generate_caption_project.mjs" `
  --video $SourceVideo --captions $Plan --out $OverlayProject `
  --interaction-state $Receipt --project-root $ProjectRoot --mode overlay

npx.cmd hyperframes check $OverlayProject --at 1 --timeout 10000 --no-contrast

$RenderMeta = Get-Content -Raw "$OverlayProject\project-meta.json" | ConvertFrom-Json
$RenderFps = "$($RenderMeta.fpsRational.num)/$($RenderMeta.fpsRational.den)"
Push-Location $OverlayProject
try {
  npx.cmd hyperframes render --format png-sequence --fps $RenderFps --output $OverlayFrames
}
finally {
  Pop-Location
}
```

Canonical overlay generation refreshes the `## Approval` section in
`captions-summary.md` with the selected style, decision actor/rationale, and
hash-binding result. It deliberately leaves rendered-frame and shared-delivery
checks pending; replace that pending line with the measured results before marking
the operation verified.

After cache deletion, rebuild the composition without replaying approval:

```powershell
node "$SkillRoot\scripts\generate_caption_project.mjs" `
  --video $SourceVideo --captions $Plan --out $OverlayProject `
  --approved-plan true --mode overlay
```

The command verifies that the frozen runtime hashes still match.

## Project Registration

Add or revise one `captions` operation in `work/project.json`. Depend on
`understanding`; also depend on the active `cut` operation when it exists.
`based_on` must equal the current dependency revisions.

```json
{
  "id": "captions",
  "skill": "video-add-captions",
  "revision": 1,
  "depends_on": ["understanding", "cut"],
  "based_on": {"understanding": 1, "cut": 1},
  "status": "verified",
  "plan": "captions/captions-plan.json",
  "outputs": ["cache/captions/overlay-frames"],
  "target": {"sequence": "main", "scope": "captions"},
  "effects": {
    "changes_timeline": false,
    "changes_geometry": false,
    "changes_video_pixels": false,
    "changes_audio": false,
    "adds_track": "captions"
  },
  "check": {"status": "pass", "report": "../review/05-captions/captions-summary.md"},
  "render": {
    "kind": "overlay",
    "asset": "cache/captions/overlay-frames",
    "asset_type": "image-sequence",
    "pattern": "frame_%06d.png",
    "start_number": 1,
    "fps": {"num": 30000, "den": 1001}
  }
}
```

Use the actual timeline FPS. Add `captions` to the active sequence operations,
then compile and render delivery once:

```powershell
python "$RepoRoot\skills\video-understand\scripts\validate.py" project `
  "$Work\project.json" $ProjectRoot
python "$RepoRoot\skills\video-understand\scripts\build_render_plan.py" $ProjectRoot
python "$RepoRoot\skills\video-understand\scripts\render_project.py" `
  "$Work\render\render-plan.json"
```

The shared renderer copies audio only for an identity timeline with no audio
filters. Cuts or retiming require encoded/filtered audio. Caption code never
overrides that decision.

## Compatibility

Without `--timeline`, `build_captions.py` still writes the old cue array.
`generate_caption_project.mjs` accepts that array with the interaction receipt.
Standalone exact ID and skip responses are legacy compatibility only.
`composite_caption_overlay.ps1` accepts an overlay video or `frame_%06d.png`
directory and writes compatible H.264/yuv420p while copying source audio. Use this
standalone path only when no active operation changes time; canonical projects use
the shared renderer.

## Self Check

Run before declaring the operation verified:

```powershell
python "$SkillRoot\scripts\check_project_protocol.py"
node "$SkillRoot\scripts\check_caption_style_config.mjs"
powershell.exe -ExecutionPolicy Bypass -File "$SkillRoot\scripts\check_structure.ps1"
node "$SkillRoot\scripts\caption_interaction.mjs" status --state $Receipt
```

Also verify:

- cue and word times are finite, ordered, within program duration, and never cross clips;
- early/middle/late captions appear on the correct mapped source pixels;
- dropped speech has no cue;
- overlay first frame exists, dimensions match, FPS equals timeline rational FPS,
  frame count equals `ceil(program_duration_s * fps.num / fps.den)`, duration covers
  the complete program, and non-caption pixels are transparent;
- the shared delivery exists, has synchronized audio, and passes the project's
  delivery report;
- `captions-summary.md` records the selected style, approval mode/rationale,
  evidence, and validation result.

HTML generation alone is not success. Inspect actual pixels and the final delivery
before reporting completion.

name: video-add-captions description: > Add word-timed captions to an Open Recut program. Use this skill to map the canonical transcript through timeline.json, review a maintained style on source-backed pixels, render a local transparent HyperFrames PNG sequence, and register it as an overlay contribution for the shared delivery render.

Video Add Captions

Dependencies

/video-understand is a prerequisite. Run it first so captions use the validated word-level 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, Python with Pillow, and Node.js >= 22 (for the .mjs scripts and npx hyperframes, fetched on demand). hyperframes render/snapshot drives a headless Chrome — it manages its own chrome-headless-shell, and falls back to a system Chrome (set CHROME to override) when the cached one is unusable. Check these before processing media.

Scope

This skill owns caption grouping, style selection, review, and the transparent caption track. It does not transcribe, cut, retime, grade, reframe, or choose the delivery audio policy.

Use video-understand first. If a cut exists, caption the active program timeline; do not treat source transcript seconds as program seconds.

Protocol Inputs

Required:

work/project.json
work/understand/transcript.json, with word-level source timestamps
work/understand/media.json
work/timeline.json
the source video named by work/project.json

The timeline is the only source-to-program mapping. All ranges are half-open [start_s, end_s). scripts/build_captions.py uses the shared projectlib.map_transcript_to_timeline; words in dropped source ranges disappear, retained words move to program time, and cues always break at clip boundaries.

Durable Outputs

work/captions/
|-- captions-plan.json
`-- caption-interaction.json
work/cache/captions/
|-- preview-project/
|-- preview-snapshots/
|-- overlay-project/
`-- overlay-frames/frame_000001.png ...
review/05-captions/
|-- captions.srt
|-- captions-style-review-<UUID>.html
|-- captions-style-review.html
|-- captions-review.html
|-- preview-early.png
|-- preview-middle.png
|-- preview-late.png
|-- preview-no-caption.png
|-- captions-evidence.json
`-- captions-summary.md

Generated HTML, copied runtime files, extracted source frames, and overlay frames are cache. The plan, SRT, decision receipt, and review evidence are durable.

Caption Plan

work/captions/captions-plan.json is the canonical caption plan, targets overlay, and uses program time. It records:

timeline_id, program_duration_s, and the source transcript path;
cue grouping settings;
approved style, decision mode, rationale, gallery choice, and complete resolved style;
source-backed review evidence and approval actor;
cues with compatibility fields start, end, text, lines, and words;
program_range and source_ranges on every cue;
program/source ranges and clip_id on every mapped word;
a local renderer recipe with exact rational FPS and hashed runtime assets.

The generator also accepts the old top-level cue array. That is a standalone compatibility adapter only; it cannot express timeline provenance or regenerate an approved protocol contribution.

Standard is the default presentation mode. Standard plans do not need a presentation field and must not be migrated merely to add one. Expressive is an explicit presentation mode, not a preset: request it with --presentation-mode expressive. The build step then adds stable cue IDs, default word semantic_role values, and a draft presentation planning shell; it does not guess layout variants.

Expressive supports only bottom-standard and center-emphasis. Read reference/caption-rules.md before filling the shell. The existing renderer consumes only completed presentation.layout_beats; it does not infer positions or change layout inside a beat.

Styles

Maintained presets are clean, minimal, social-bold, pill, boxed, stroked, and shorts. Gallery combination IDs add maintained themes, for example pill-yellow, stroked-blue, and shorts-green. Karaoke is an option, not a separate preset.

Default delegated choice: clean, Karaoke off. Choose a more expressive preset only when the destination or footage justifies it. Store job-specific changes in an overrides JSON; do not edit the generator for ordinary style feedback.

Expressive supports both semantic-only (Karaoke off) and semantic-plus-karaoke (Karaoke on). Semantic emphasis persists for the full cue in both configurations. In coexistence mode, semantic emphasis and the active Karaoke word use style.wordHighlight.activeColor as their foreground color in every highlight mode; a background highlight remains an additional effect. The effective scale is max(semantic scale, karaoke active scale), never their product. Expressive never renders an underline, including for the compatible contrast semantic role.

The existing review contract continues to use semantic-only as the primary Expressive evidence and binds semantic-plus-karaoke separately. The compatibility payload field remains named experimental_comparison; do not rename it without an approval-protocol migration.

Read:

reference/caption-rules.md before changing grouping limits;
reference/caption-style-themes.md for style vocabulary;
reference/caption-feedback-mapping.md for override mapping.

Decision Modes

Two honest modes share the same hash-bound state machine:

human: record the user's exact gallery response with select, show all real preview evidence, and record the exact confirmation response with confirm.
agent: use only when the user explicitly delegates caption decisions. Start with a non-empty delegation note, then use agent-select and agent-confirm with non-empty rationales.

Human commands fail in agent mode and agent commands fail in human mode. Both modes require four existing source-backed images before approval. A change to the source, plan, timeline, style, override, project metadata, review page, or evidence invalidates approval; rebuild the affected artifacts and review them again.

Canonical Workflow

Run from the repository root. Set project paths:

$RepoRoot = (Resolve-Path ".").Path
$SkillRoot = Join-Path $RepoRoot "skills\video-add-captions"
$ProjectRoot = (Resolve-Path "<project-root>").Path
$Work = Join-Path $ProjectRoot "work"
$SourceVideo = "<absolute path from work/project.json>"
$Plan = Join-Path $Work "captions\captions-plan.json"
$Receipt = Join-Path $Work "captions\caption-interaction.json"
$Review = Join-Path $ProjectRoot "review\05-captions"
$Cache = Join-Path $Work "cache\captions"
$PreviewProject = Join-Path $Cache "preview-project"
$ComparisonProject = Join-Path $Cache "preview-project-expressive-karaoke"
$OverlayProject = Join-Path $Cache "overlay-project"
$OverlayFrames = Join-Path $Cache "overlay-frames"
New-Item -ItemType Directory -Force -Path (Split-Path $Plan),$Review,$Cache | Out-Null

Before building the caption plan, ask the user to choose the presentation mode unless the current request already explicitly selects one. Render the question in the user's current conversation language while keeping the Standard and Expressive identifiers untranslated. Keep the English source prompt below in this file, present the localized choice in the current session, then Present + STOP:

Choose the caption presentation mode:

1. Standard: Uses stable conventional placement and suits most content.
2. Expressive: Switches between bottom and center placement based on meaning and emphasizes keywords and numbers.

Reply with Standard, Expressive, or "Use the default (Standard)."

Do not run build_captions.py until the user responds. If the user explicitly delegates the choice in any wording, select Standard. State that Standard was selected as the compatibility-preserving default before building the plan.

Build program-time cues and the review SRT:

python "$SkillRoot\scripts\build_captions.py" `
  "$Work\understand\transcript.json" `
  $Plan `
  "$Review\captions.srt" `
  --timeline "$Work\timeline.json" `
  --source-transcript "understand/transcript.json" `
  --max-chars 42 --max-lines 2 --max-dur 6 --gap 0.6

After the user selects Standard, the command above produces the unchanged Standard plan. After the user selects Expressive, run the same command with:

  --presentation-mode expressive

Before starting style selection or generating a preview, the Agent must plan the entire Expressive program in one pass:

Read the transcript, available understanding artifacts, canonical timeline, generated caption cues, and any real source-frame evidence needed to justify layout changes.
Fill all presentation.layout_beats, annotate exceptional word semantic_role values, write one rationale per beat, and write the overall presentation.planner.rationale.
Set presentation.planning_status to complete only after every cue is covered exactly once.
Validate the completed plan before preview generation:

python "$SkillRoot\scripts\build_captions.py" --validate-plan $Plan

Do not ask the user to choose a position cue by cue. The user reviews the existing HTML evidence and corrects only a small number of anomalous beats. Starting the interaction after planning keeps the existing caption-plan hash binding authoritative.

Start one decision mode. For human mode, publish the maintained offline gallery into the project review directory without asking the user to locate a file:

$StartOutput = node "$SkillRoot\scripts\caption_interaction.mjs" start `
  --state $Receipt --source $SourceVideo --captions $Plan `
  --review-dir $Review --no-open true
$StartOutput | Write-Host

start prints the authoritative captions-style-review-<UUID>.html project page path and the fixed question. captions-style-review.html is a non-authoritative latest convenience alias only. The Agent must take the UUID path printed after Caption style review:, assign that exact path to $StyleReviewPage, and open it with the native command for the host OS:

Windows PowerShell:

$StyleReviewPage = "<authoritative UUID page path printed after Caption style review:>"
Start-Process -FilePath (Resolve-Path $StyleReviewPage)

macOS:

open "$StyleReviewPage"

Linux:

xdg-open "$StyleReviewPage"

If opening fails, diagnose the command or path and retry it. Do not ask the user to find the page. Present the opened gallery and the fixed question exactly as printed, then Present + STOP. Do not continue until the human copies the structured summary from the page. It has this form:

Caption style review
Review: <UUID from the opened page>
Decision: select
Choice: pill-yellow

Pass the user's exact summary unchanged to select --response. In PowerShell, a single-quoted here-string preserves the lines safely:

$StyleResponse = @'
Caption style review
Review: <UUID from the opened page>
Decision: select
Choice: pill-yellow
'@
node "$SkillRoot\scripts\caption_interaction.mjs" select `
  --state $Receipt --response $StyleResponse

The block above shows the structure; use the user's returned block, including their review UUID and choice. Never infer, normalize, or simulate a human response.

For explicit delegation, create a separate Agent-mode receipt. Keep the bound gallery, suppress the script's Windows-only auto-open, open the authoritative UUID page with the matching native command above, and inspect it before choosing:

$StartOutput = node "$SkillRoot\scripts\caption_interaction.mjs" start `
  --state $Receipt --source $SourceVideo --captions $Plan `
  --review-dir $Review `
  --decision-mode agent `
  --delegation-note "User delegated caption style and preview approval." `
  --no-open true
$StartOutput | Write-Host

node "$SkillRoot\scripts\caption_interaction.mjs" agent-select `
  --state $Receipt --choice clean `
  --rationale "Conservative readable treatment preserves the talking-head frame."

Use agent-select only after inspecting the bound gallery and record the real rationale. Do not manufacture a human summary or use select in Agent mode.

Generate a transparent preview composition. The project contains only local GSAP and font files and preserves rational FPS in project-meta.json:

node "$SkillRoot\scripts\generate_caption_project.mjs" `
  --video $SourceVideo --captions $Plan --out $PreviewProject `
  --interaction-state $Receipt --project-root $ProjectRoot --mode preview

npx.cmd hyperframes check $PreviewProject --at 1 --timeout 10000 --no-contrast

For the primary Expressive review, append --karaoke false. Generate the supported coexistence comparison from the same source, plan, interaction selection, and timing by changing only the output directory and appending --karaoke true:

node "$SkillRoot\scripts\generate_caption_project.mjs" `
  --video $SourceVideo --captions $Plan --out $ComparisonProject `
  --interaction-state $Receipt --project-root $ProjectRoot `
  --karaoke true --mode preview

For Standard, build mapped early/middle/late/no-caption evidence. This command captures transparent HyperFrames snapshots, maps each program time through timeline.json, extracts the matching original source frame, and composites the two with Pillow:

python "$SkillRoot\scripts\build_caption_review.py" `
  --source $SourceVideo --timeline "$Work\timeline.json" --plan $Plan `
  --project $PreviewProject --cache "$Cache\review-cache" --out $Review `
  --interaction-state $Receipt

For Expressive, the same builder emits one primary midpoint sample per layout beat, one no-caption sample, and a separately bound semantic-only/Karaoke comparison:

python "$SkillRoot\scripts\build_caption_review.py" `
  --source $SourceVideo --timeline "$Work\timeline.json" --plan $Plan `
  --project $PreviewProject --comparison-project $ComparisonProject `
  --cache "$Cache\review-cache" --out $Review --interaction-state $Receipt

The builder writes captions-review.html as the one authoritative page. Standard contains exactly the four maintained evidence labels. Expressive contains the dynamic layout beat evidence, filter controls, image enlargement, no-caption evidence, and the separately bound coexistence comparison outside the primary approval count. The Agent must open that page with the native command for the host OS; for example, set $EvidenceReviewPage = "$Review\captions-review.html" and run one of:

Start-Process -FilePath (Resolve-Path $EvidenceReviewPage)

open "$EvidenceReviewPage"

xdg-open "$EvidenceReviewPage"

If opening fails, diagnose and retry. Inspect every primary image for readability, safe-area placement, clipping, word wrapping, stable beat placement, and unwanted pixels in preview-no-caption.png. Standard binds the maintained four images:

$Evidence = @(
  "$Review\preview-early.png",
  "$Review\preview-middle.png",
  "$Review\preview-late.png",
  "$Review\preview-no-caption.png"
) -join ","

node "$SkillRoot\scripts\caption_interaction.mjs" preview-ready `
  --state $Receipt --project-meta "$PreviewProject\project-meta.json" `
  --evidence $Evidence --review-page "$Review\captions-review.html" `
  --timeline "$Work\timeline.json"

Expressive binds the dynamic primary count and the separate coexistence pair from captions-evidence.json:

$EvidenceDocument = Get-Content "$Review\captions-evidence.json" -Raw | ConvertFrom-Json
$Evidence = @($EvidenceDocument.samples | ForEach-Object {
  Join-Path $Review $_.preview
}) -join ","
$ComparisonEvidence = @($EvidenceDocument.experimental_comparison.samples | ForEach-Object {
  Join-Path $Review $_.preview
}) -join ","

node "$SkillRoot\scripts\caption_interaction.mjs" preview-ready `
  --state $Receipt --project-meta "$PreviewProject\project-meta.json" `
  --evidence $Evidence --comparison-evidence $ComparisonEvidence `
  --review-page "$Review\captions-review.html" --timeline "$Work\timeline.json"

In human mode, present the opened page and its images, then Present + STOP. The human copies one of these structured summaries from the page:

Caption preview review
Review: <UUID from the opened page>
Decision: approve
Evidence: early, middle, late, no-caption

Expressive approval uses the dynamic evidence token and requires the explicit Karaoke choice made after inspecting the bound comparison evidence:

Caption preview review
Review: <UUID from the opened page>
Decision: approve
Evidence: expressive-layout-beats
Karaoke: on|off

Caption preview review
Review: <UUID from the opened page>
Decision: revise
Changes: Raise captions 20 pixels.

Pass the returned summary unchanged to confirm --response or adjust --response. Quote the multiline response safely in PowerShell:

$PreviewResponse = @'
Caption preview review
Review: <UUID from the opened page>
Decision: approve
Evidence: early, middle, late, no-caption
'@
node "$SkillRoot\scripts\caption_interaction.mjs" confirm `
  --state $Receipt --response $PreviewResponse

For a revise summary, preserve it in $PreviewResponse the same way and run:

node "$SkillRoot\scripts\caption_interaction.mjs" adjust `
  --state $Receipt --response $PreviewResponse

Apply the requested change, then regenerate the preview project, review page, and all evidence before presenting the gate again. Never convert human feedback into an Agent decision.

In delegated mode, the Agent must inspect the same captions-review.html page and all primary evidence plus any coexistence comparison before recording its own rationale. For Standard, keep the existing command unchanged:

node "$SkillRoot\scripts\caption_interaction.mjs" agent-confirm `
  --state $Receipt `
  --rationale "All bound primary frames are readable and collision-free; coexistence comparison evidence was inspected separately when present."

For Expressive, pass the explicit Karaoke choice made from the comparison:

node "$SkillRoot\scripts\caption_interaction.mjs" agent-confirm `
  --state $Receipt --karaoke off `
  --rationale "The semantic-only comparison is clearer for this footage."

Generate the approved overlay project. This formal render is the only step that finalizes style, review, and hashed runtime assets in the canonical plan. Expressive overlay generation uses the preview-approved Karaoke choice; an explicit --karaoke value must match that approval:

node "$SkillRoot\scripts\generate_caption_project.mjs" `
  --video $SourceVideo --captions $Plan --out $OverlayProject `
  --interaction-state $Receipt --project-root $ProjectRoot --mode overlay

npx.cmd hyperframes check $OverlayProject --at 1 --timeout 10000 --no-contrast

$RenderMeta = Get-Content -Raw "$OverlayProject\project-meta.json" | ConvertFrom-Json
$RenderFps = "$($RenderMeta.fpsRational.num)/$($RenderMeta.fpsRational.den)"
Push-Location $OverlayProject
try {
  npx.cmd hyperframes render --format png-sequence --fps $RenderFps --output $OverlayFrames
}
finally {
  Pop-Location
}

Canonical overlay generation refreshes the ## Approval section in captions-summary.md with the selected style, decision actor/rationale, and hash-binding result. It deliberately leaves rendered-frame and shared-delivery checks pending; replace that pending line with the measured results before marking the operation verified.

After cache deletion, rebuild the composition without replaying approval:

node "$SkillRoot\scripts\generate_caption_project.mjs" `
  --video $SourceVideo --captions $Plan --out $OverlayProject `
  --approved-plan true --mode overlay

The command verifies that the frozen runtime hashes still match.

Project Registration

Add or revise one captions operation in work/project.json. Depend on understanding; also depend on the active cut operation when it exists. based_on must equal the current dependency revisions.

{
  "id": "captions",
  "skill": "video-add-captions",
  "revision": 1,
  "depends_on": ["understanding", "cut"],
  "based_on": {"understanding": 1, "cut": 1},
  "status": "verified",
  "plan": "captions/captions-plan.json",
  "outputs": ["cache/captions/overlay-frames"],
  "target": {"sequence": "main", "scope": "captions"},
  "effects": {
    "changes_timeline": false,
    "changes_geometry": false,
    "changes_video_pixels": false,
    "changes_audio": false,
    "adds_track": "captions"
  },
  "check": {"status": "pass", "report": "../review/05-captions/captions-summary.md"},
  "render": {
    "kind": "overlay",
    "asset": "cache/captions/overlay-frames",
    "asset_type": "image-sequence",
    "pattern": "frame_%06d.png",
    "start_number": 1,
    "fps": {"num": 30000, "den": 1001}
  }
}

Use the actual timeline FPS. Add captions to the active sequence operations, then compile and render delivery once:

python "$RepoRoot\skills\video-understand\scripts\validate.py" project `
  "$Work\project.json" $ProjectRoot
python "$RepoRoot\skills\video-understand\scripts\build_render_plan.py" $ProjectRoot
python "$RepoRoot\skills\video-understand\scripts\render_project.py" `
  "$Work\render\render-plan.json"

The shared renderer copies audio only for an identity timeline with no audio filters. Cuts or retiming require encoded/filtered audio. Caption code never overrides that decision.

Compatibility

Without --timeline, build_captions.py still writes the old cue array. generate_caption_project.mjs accepts that array with the interaction receipt. Standalone exact ID and skip responses are legacy compatibility only. composite_caption_overlay.ps1 accepts an overlay video or frame_%06d.png directory and writes compatible H.264/yuv420p while copying source audio. Use this standalone path only when no active operation changes time; canonical projects use the shared renderer.

Self Check

Run before declaring the operation verified:

python "$SkillRoot\scripts\check_project_protocol.py"
node "$SkillRoot\scripts\check_caption_style_config.mjs"
powershell.exe -ExecutionPolicy Bypass -File "$SkillRoot\scripts\check_structure.ps1"
node "$SkillRoot\scripts\caption_interaction.mjs" status --state $Receipt

Also verify:

cue and word times are finite, ordered, within program duration, and never cross clips;
early/middle/late captions appear on the correct mapped source pixels;
dropped speech has no cue;
overlay first frame exists, dimensions match, FPS equals timeline rational FPS, frame count equals ceil(program_duration_s * fps.num / fps.den), duration covers the complete program, and non-caption pixels are transparent;
the shared delivery exists, has synchronized audio, and passes the project's delivery report;
captions-summary.md records the selected style, approval mode/rationale, evidence, and validation result.

HTML generation alone is not success. Inspect actual pixels and the final delivery before reporting completion.