agentleFS
first-flukefirst-fluke/fullstack-starter/.agents/skills/oma-video/SKILL.md

oma-video

Skill223 starsChanged 28 days agoMITView on GitHub

Create short, explainer, or recorded-demo videos through the OMA video CLI. Use for scripts, narration, assets, composition, and video delivery.

We split this skill into the rules it gives an agent

8 rules. About 0 are common, 4 are distinctive, and 4 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈4

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 4

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Use this skill for a short/reel, README or code explainer, or demo walkthrough.

    line 12
  2. Very specific:

    Use oma-image for a still image, oma-slide for a deck, oma-voice for audio only, and oma-explanation for an interactive HTML explainer.

    line 22
  3. Distinctive:

    Ask only when it changes the result: an ambiguous mode/source, a required demo recording, or a cost confirmation.

    line 53
  4. Very specific:

    Keep output and capture paths inside $PWD unless external output is explicitly allowed.

    line 65
  5. Very specific:

    Before a paid provider action or rerun, compare the planning estimate with cost.guardrail_usd or --max-usd and reuse existing spend …

    line 72
  6. Distinctive:

    Keep run directories. Never auto-prune a user’s video artifacts.

    line 80
  7. Very specific:

    Read video and asset paths from the manifest.

    line 109
  8. Distinctive:

    Do not run a live render merely to test documentation or a mock-only branch.

    line 159

≈ “Common” is an estimate: at least 1 in 5 of the 308,232 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Video Router

Scheduling

When to use

Use this skill for a short/reel, README or code explainer, or demo walkthrough.

RequestModeDefault aspectRequired source
Short, reel, social clipshorts9:16Topic or brief
README, code, data explanationexplainer16:9Topic or source path
Demo or walkthroughdemo16:9Human recording via --capture

When NOT to use

Use oma-image for a still image, oma-slide for a deck, oma-voice for audio only, and oma-explanation for an interactive HTML explainer. Editing an existing finished video and live streaming are out of scope.

Structural Flow

Inputs are a brief plus optional mode, aspect, locale, captions, visual, voice, music, duration, compositor, capture path, and seed. Outputs live in .agents/results/videos/<timestamp>-<shortid>-<mode>/:

  • script.json, timing.json, and render-spec.json form the deterministic asset bus.
  • Captions and acquired audio/visual assets are recorded in manifest.json with hashes, providers, cost, warnings, and exit code.
  • A successful real render contains <mode>-<slug>.mp4, an encoded video stream, and a positive ffprobe duration.

OMA_VIDEO_MOCK=1 is a test harness only. It can create deterministic text placeholders with an .mp4 name; those files are never a user deliverable. A missing HyperFrames/MPT toolchain, an un-authored composition, a render error, or an invalid video fails with diagnostics and leaves the script/render spec for recovery.

Decide and confirm

Infer mode when clear: short/reel -> shorts; README/code/data/explain -> explainer; demo/walkthrough/capture -> demo. For a one-line request, state the inferred mode, aspect, duration, visual strategy, captions, locale, and voice/music before invoking. Do not make the user complete a questionnaire when those defaults are clear.

Ask only when it changes the result: an ambiguous mode/source, a required demo recording, or a cost confirmation. Respect an explicit mode, aspect, duration, captions, or voice verbatim.

For a demo, a human records the screen and controls login. --source web --url provides context only; it never automates login or starts a recorder. Without --capture, return guided capture instructions and stop.

Logical Operations

Guardrails

  1. Keep output and capture paths inside $PWD unless external output is explicitly allowed. Validate capture formats and copy external assets into the run directory. Mask URL query/hash tokens in logs and manifests.
  2. Provider configuration is key-optional: use the configured chain. Paid providers require their environment key and the cost guardrail. Local fallbacks may replace voice, visuals, captions, or music; record coverage in warnings. A compositor failure is never a fallback video.
  3. Before a paid provider action or rerun, compare the planning estimate with cost.guardrail_usd or --max-usd and reuse existing spend authorization. In an active OMA video workflow, record the actual paid, limited, fallback, or declined choice before executing it (execution protocol Step 2). Pass --yes or OMA_VIDEO_YES=1 only for an already authorized paid action; the event does not grant permission.
  4. Respect limits.max_duration_sec (180) and limits.max_scenes (40). Cancel subprocess work on SIGINT/SIGTERM.
  5. Keep run directories. Never auto-prune a user’s video artifacts.
  6. --dry-run writes only planning artifacts and does no provider render. It does not prove an MP4 exists.

Canonical command path

# Plan or create the asset bus. Supply --script whenever an agent authored it.
oma video generate "Jeju coffee" --mode shorts --aspect 9:16 \
  --captions tiktok --script ./script.json --output json

# Deterministic planning only; no real render or provider work.
oma video generate "explain this project" --mode explainer --seed 42 --dry-run

# Human-recorded demo input.
oma video generate "feature walkthrough" --mode demo --capture <absolute-path>.mp4

# Scaffold the per-run HyperFrames project, author index.html as instructed,
# then render and validate the encoded output.
oma video compose <runDir> --output json
oma video render <runDir> --output json

# Diagnose required toolchains without changing a run.
oma video doctor
oma video provider list

oma video generate --output json returns {exitCode, runDir, manifestPath, scriptPath, renderSpecPath, warnings, error}. Read video and asset paths from the manifest; the JSON envelope has no outputs field.

Failure and recovery

  • Missing HyperFrames composition: oma video compose <runDir> prepares the project and authoring contract. Author <runDir>/hyperframes/index.html using the generated AUTHORING.md, then invoke oma video render <runDir>. The command lints, renders, and ffprobes the output. Fix a reported composition or toolchain failure and re-run; return a failure report when it cannot render.
  • Missing MPT toolchain: --compositor mpt requires the installed checkout, its virtual environment, and ffmpeg. Use oma video doctor --install-mpt when setup is authorized and available. MPT setup failures, driver failures, and non-video output fail with the diagnostic; they do not write a placeholder MP4.

Success means all asset schemas and manifest hashes validate and a real MP4 passes video-stream and duration validation. A partial success may use a key-free visual, timing, caption, or music fallback, but it still requires that real video validation.

References

Conditional resources

Load only what the task needs:

  • resources/execution-protocol.md for the full ordered pipeline, failure mapping, and JSON reporting rules.
  • resources/vendor-matrix.md before changing providers, keys, cost, or fallback order.
  • resources/script-schema.md when authoring or validating --script input.
  • resources/prompt-tips.md when turning a brief into scene prompts.
  • resources/hyperframes-authoring/README.md and the selected mode guide before writing index.html.
  • resources/checklist.md before handing a real video to a user.

Verification

For CLI/runtime changes, add a regression test for the affected success and failure paths. At minimum run the focused Vitest files, for example:

cd cli
bunx vitest run commands/video/providers/compositor.test.ts \
  commands/video/orchestrator.test.ts
bunx biome check commands/video/providers/compositor.ts \
  commands/video/providers/compositor.test.ts

Do not run a live render merely to test documentation or a mock-only branch.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

Reviews can't be read right now.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.