- Home
- Skills
- Image & Video
- Video Orchestrator: Multi-Round AI Video Builds
More screenshots
Works with the AI Agents you already use
Video Orchestrator: Multi-Round AI Video Builds
Run AI video builds that exceed one generation round: a durable scene queue in a JSON state file, chain-continue tracking, per-clip QC gates, a failure policy, and CPU post scheduling.
$5
Video Orchestrator: Multi-Round AI Video Builds
Example session with this skill installed
I'm building a 20-scene music video but my generator only allows 6 clips per message, and I keep losing track of what's done when I stop and come back. Some scenes continue from the previous scene's last frame. Set up the build so it survives interruptions.
- Read your context and instructions
- Compiled the video orchestrator
BUILD SETUP — produced by the skill's state contract and estimate.
The estimate, before anything generates: rounds = max(ceil(20/6), longest chain). A 5-scene chain makes the floor 5 rounds, not 4 — chains serialize. Generation budget: 20 scenes + retry allowance (worst case one retry each) = plan for up to 40 calls, and say so now. Post time is estimated separately: interpolation for the fast-motion scenes runs as an overnight CPU batch; still upscaling is priced per hero frame, never per clip.
The state file, work/queue.json, created before round 1: one record per scene with id, time slot (slots tile the track exactly — we verify the 20 slots sum to the track length with no gaps or overlaps before starting), prompt, status queued, attempts 0, clip path null, the generator request record slot, and for chained scenes the parent id plus a snapshot field. A project section lists post targets: which scenes get interpolation, which frames get upscaled.
Round 1 runs the protocol: pick scenes in timeline order up to 6, skipping any chained scene whose parent is not done yet. Each chained scene extracts its parent's last frame fresh at generation time — never a stale PNG. Every clip gets QC'd on arrival (decodes cleanly, duration and resolution match the batch standard, middle frame checked against the brief, first frame not black), the state file is updated after each clip, and the request record is saved immediately.
Interruption drill, because it will happen: you stop after clip 4 of round 2. Nothing is lost — the state file is current to clip 4. The picker only looks at scenes still marked queued, so resuming is literally running the protocol again. Scenes are never deleted, only re-marked, so the final report shows what was planned, attempted, and substituted.
Failure handling, pre-agreed: a clip that fails QC gets one retry with a simplified prompt. A second failure marks it failed, and its slot is filled by extending an adjacent done scene or a neutral filler — recorded in the round log. If a chain parent fails, its children are re-pointed at the nearest done scene or marked skipped with a note. No silent orphans, no holes at assembly.
Connects securely to your tools. The creator never sees your data.
About this skill
AI video generators cap how many clips you can request per message, so any real project — a music video, a multi-scene explainer — spans several rounds and usually several interruptions. Builds run from memory or chat history die at the first interruption: nobody knows which scenes are done, which failed, or what a chained shot was supposed to continue from. This skill replaces memory with machinery. All state lives in one queue.json file, updated after every clip, never at the end of the round — a crash costs at most one clip's bookkeeping. Each scene record carries its timeline slot (slots must tile the track exactly), prompt, status, attempt counter, clip path, the generator's request record, and chain links with snapshot handles. The round protocol picks up to the cap in timeline order, respects chain dependencies without exception, extracts each chained scene's parent last frame fresh, and QCs every clip — decode check, duration, resolution, middle frame against the brief, first frame not black. The failure policy is explicit: retry once with a simpler prompt, then mark failed and fill the slot with a documented fallback; downgrades are announced, never silent. A pre-build estimate does the honest arithmetic first — rounds are floored by the longest chain, generation budget includes retries, and post-production time is estimated separately. Post spends CPU instead of generations: frame interpolation only for fast-motion scenes, neural upscaling for hero stills only, because full-clip upscaling on CPU runs hours per minute of footage. A worked example runs a 12-scene build under a 6-per-message cap, and a build-health scorecard keeps every round honest.
Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Video Skills.
- It orchestrates generation; it does not write the creative plan. Scene prompts, slots, and the track come from your planning workflow.
- Schedules assume the cap and generator behavior stay roughly constant; a mid-build cap change adapts the picker but stretches the estimate.
- CPU post-production timings are hardware-dependent — interpolation and upscaling estimates are planning figures to measure against your first batch, not guarantees.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before buying a skill.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed 24 days ago
Permissions
No special permissions declared or detected