<< All versions

Skill v1.0.0

currentAutomated scan100/100
galbaz1/video-research-mcp/video-generation
──Details
PublishedSeptember 26, 2026 at 06:52 AM
Content Hashsha256:42d188fd62841fed...
Git SHA
──Files
Files (1 file, 3.1 KB)
SKILL.md3.1 KBactive
SKILL.md · 30 lines · 3.1 KB

version: "1.0.0" name: video-generation description: Generate AI video clips with an available provider, including image-to-video, reference assets, or extensions. Use for generation requests, not video analysis or FFmpeg editing.


Video Generation

This plugin packages production guidance; it does not bundle a Veo MCP server or a video-generation script. Discover the connected provider and read its current tool schema before calling it. If none is available, preserve a generation-ready brief and identify the missing provider.

Workflow

  1. Fix the brief. Record subject, motion, camera, duration, aspect ratio, audio, destination, and supplied assets. Use an anchor image when continuity matters; text-only generation is also supported where the provider permits it. Completion: one bounded shot specification with explicit acceptance criteria.
  2. Verify capability. Read provider-details.md for the current provider snapshot, then check live model availability, input constraints, pricing, and the actual tool/API schema. Do not translate a provider feature into an invented MCP tool name. Completion: a supported request and an authorized spend/attempt limit.
  3. Generate one draft. Keep its prompt, model, settings, reference paths, operation ID, and output together. A running operation is pending; poll that operation rather than launching a duplicate. Completion: terminal result or preserved provider failure.
  4. Inspect the artifact. Download the successful output, probe duration/resolution/audio with ffprobe, inspect a contact sheet and motion playback. Check subject identity, object integrity, timing, and audio against the brief. Completion: evidence for every acceptance criterion, including failures.
  5. Repair only a measured defect. Change one variable, retain the earlier output, and use at most one repair attempt within the existing budget. Stop on a repeated infrastructure failure, exhausted budget, or terminal refusal. Completion: an accepted local clip or a precise unresolved defect.

For an approved draft, generate a final only if the requested delivery settings require it. Re-check the final artifact; a winning draft does not prove the final has the same quality. For multi-shot continuity and assembly, use video-production; for encoding use ffmpeg-production.

Prompt Template

[Subject and setting]. [One primary action over the requested duration]. Camera [movement and framing]. Light comes from [physical source]. Maintain [specific reference features]. Audio: [dialogue, ambience, or intended silence]. Deliver [aspect ratio and supported resolution].

Include exact spoken words when dialogue matters. Preserve the user's exclusions using a supported negative-prompt field or clear instructions. Reference images guide appearance; they do not guarantee identity or text fidelity.

Delivery

Return the playable clip path, inspected properties, provider/model, reference provenance, and remaining defects. Distinguish prepared, pending, failed, and verified local output. Publication requires the user's destination authority.

All versions