Miles Video Editing

Miles Video Editing

Turn one real source video into one semantic, reviewable candidate. The visual system must explain the speech; repeating subtitles in decorative boxes is not completion.

Use the public local toolchain

This Miles Skill owns the V10 method, portable workspace, storyboard contract, composition builder, spatial rules, and acceptance. A user needs only this plugin plus public local system prerequisites; they do not need another Skill, Miles API, Miles credentials, or a Miles-hosted service.

Use the host-provided active Skill catalog as the only availability authority when the user explicitly delegates an operation to another Skill. Keep the catalog decision and keep the selected executor unchanged. Do not rediscover or replace the selected executor from disk folders, caches, inventories, or configuration. The default path does not delegate: it installs the pinned Apache-2.0 hyperframes@0.7.81 npm package inside the current video project and forces local Whisper transcription.

Never infer a transcript, media asset, successful check, or successful render.

Read [the V10 contract](references/v10-contract.md) before authoring the storyboard and [the dependency contract](references/dependencies.md) before any setup or render operation.

Create an isolated project

Require one readable video with a video stream and audible primary speech. Run the non-mutating environment check first:

python3 /scripts/video_workspace.py preflight --input  --json

Preflight never installs, downloads, or renders. A blocker stops the workflow. When ready, initialize a new project outside the Skill directory:

python3 /scripts/video_workspace.py init \
  --input  --project-dir 

The source is copied into source/ and never overwritten. All manifest paths are relative and must resolve inside the project.

With explicit approval for network and disk changes, install the project-local public toolchain and pinned browser:

python3 /scripts/video_workspace.py setup \
  --project-dir 

Setup installs only into work/toolchain/. It never reads an API key or writes to the Skill package.

Obtain evidence and author the storyboard

Transcribe the copied source locally:

python3 /scripts/video_workspace.py transcribe \
  --project-dir  --language zh

This forces local Whisper and stores SRT at work/transcript.srt. Inspect and correct the result against the real audio before authoring semantic beats; do not publish the transcript with this Skill.

Create work/spec/storyboard.json from [the public schema](references/storyboard.schema.json). Each semantic beat must state the spoken claim it explains, its time range, one visual treatment, and why that treatment adds information. Mark required real media as missing until actually supplied or reconstructed from verified evidence.

Validate before building:

python3 /scripts/video_workspace.py validate \
  --project-dir 

Missing media, transcript mismatch, out-of-range timing, duplicate IDs, same-lane overlap, or unsafe lower-screen overlays block the build.

Build and check the composition

Generate the portable HyperFrames project:

python3 /scripts/video_workspace.py build \
  --project-dir 

The builder uses deterministic finite WAAPI animation and does not bundle GSAP, private fonts, remote scripts, Miles media, or machine-specific paths.

From work/composition/, use the project-local pinned CLI:

npm install --ignore-scripts --no-audit --no-fund
npx hyperframes check --strict --snapshots
npx hyperframes preview

Package, browser, font, and model setup may use the public npm, Google Fonts, or Hugging Face endpoints; perform them only after explicit user approval. None is a Miles API. A passing check is not render approval. Show the final Studio preview and wait for approval.

Render and verify

After preview approval:

npx hyperframes render --quality high --output ../../outputs/candidate.mp4
python3 /scripts/video_workspace.py verify \
  --project-dir  --output outputs/candidate.mp4 --json

Verify a real non-empty H.264/AAC 1080x1920 MP4, duration, frame rate, audio, and snapshots. Call it a candidate until semantic timing, explanatory value, caption readability, and spatial crowding pass human review. Automated checks cannot promote this Skill from candidate to released-owned.

Protect Miles personal information

Allow public Miles branding, but remove non-brand contact details, private paths, account identifiers, private chats, and unnecessary face, voice, device, location, or source-media metadata from previews, renders, logs, and handoffs. Keep required private source media local and stop before any unapproved upload or public export.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: Add-Miles
• Source: Add-Miles/mileswang-skill
• License: MIT

Install and usage instructions live in the source repository linked above.