Plotloom Video Adapter

When to Use

Use when episodes/epXXX/video-prompts.md or video-prompts-en.md exists and a clip candidate should be produced or queried.

Inputs

• video-prompts.md / video-prompts-en.md
• Target clip id such as clip-01
• Adapter choice: fake adapter by default; Dreamina only for explicit real generation.

Outputs

• episodes/epXXX/videos/clip-YY/candidates/vNNN.mp4
• If queueing, a human-readable queue note near the clip folder containing submit_id, status, adapter, and next query command.

Read These Resources When...

• Read references/fake-adapter.md for contract tests and no-quota dry runs.
• Read references/dreamina-cli.md before any real Dreamina/即梦 submission or polling.
• Use templates/adapter-request.md to record preflight, command, status, and handoff.

Workflow

1. Read the prompt file and target clip id.
2. Choose fake adapter by default for contract tests.
3. Use Dreamina only when explicitly running real generation and preflight passes.
4. Submit one candidate at a time.
5. Save outputs to episodes/epXXX/videos/clip-YY/candidates/vNNN.mp4.
6. If the provider queues, preserve submit_id in a visible note near the clip folder, not a hidden runtime DB.
7. Hand off to plotloom-asset-selection after a candidate exists.

Adapter Rules

• Fail fast on missing prompt, missing clip id, invalid output path, missing login, missing maestro, missing quota, or provider error.
• Fake adapter proves file-path contracts and ffmpeg compatibility; it does not prove creative quality.
• Dreamina requires host pre-authentication. The adapter must not automate OAuth, store credentials, or copy tokens.
• Do not create a Python runtime client in the skill; keep this as a prompt and command contract plus the plotloom CLI.
• Do not batch-generate three candidates in MVP.

Reference Discipline

Reference handling is high risk. Do not assume that a local path written inside prompt prose is sent to the provider.

Before any image-to-video or reference-to-video submission:

1. Identify the exact intended first frame and reference images for this clip only.
2. Verify those files or asset://... ids are the actual request inputs, not just Markdown text.
3. Reject ambiguous phrasing such as Use Image 1 unless the Image mapping has been programmatically resolved for the target clip.
4. Record the submitted reference list in a visible receipt/handoff. If the CLI cannot show the actual refs, stop and add/patch the tool support before expensive generation.

For VolcEngine/Seedance identity work:

• Full character sheets may trigger InputImageSensitiveContentDetected.PrivacyInformation.
• Transparent face mesh, red topology, blur, or privacy mesh are not reliable bypasses.
• Prefer provider-approved asset://... virtual face anchors for faces, plus face-blocked costume/body sheets for clothing and silhouette.
• Treat asset://... as face-only. Hats, wardrobe, body type, props, and scene blocking still need text or non-sensitive visual refs.
• Face consistency smoke tests should start with medium close-up, front-left 3/4 face, visible eyes, no deep hat shadow, face occupying about 25-35% of frame, and minimal action.

Batch Safety

For repeated candidate or asset generation, use a manifest/resume pattern: skip outputs that already exist unless --force is explicit, record per-item status, and continue after timeouts. Do not restart a batch from item 1 after one successful expensive generation.

Stop Conditions

Stop after candidate creation, queue note, or provider failure report. Do not select or stitch.

Next Skill Handoff

Use plotloom-asset-selection to review and accept/reroll the candidate.

Failure Modes

• Fake adapter failure: check ffmpeg installation and output path.
• Dreamina not logged in: run dreamina user_credit manually on the host.
• Dreamina account not maestro: stop and report permission gap.
• Queueing: record submit_id, status, and query command visibly.
• Generation failed: preserve error and suggest prompt revision only if failure is prompt-related.