Menu
Create, run, resume, or test StoryTeller AI video pipelines.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
Create the AI video preproduction package.
Write keyframe image requests and a video prompt for one shot.
Write one concept image request and local concept manifest.
Based on SOC occupation classification
| name | ai-video-pipeline |
| description | Create, run, resume, or test StoryTeller AI video pipelines. |
| Stage | Execute | Done when |
|---|---|---|
director | create subagent: $ai-video-director | Director validator passes. |
concept_requests | create one subagent per pending concept: $ai-video-concept-art | Concept-art validator passes; status.py sees no pending concept requests. |
concept_images | create one subagent per pending concept image request: $imagegen | Each request's output_path exists. |
shot_requests | create one subagent per pending shot: $ai-video-shot-producer | Shot-producer validator passes; status.py sees no pending shot requests. |
keyframe_images | create one subagent per pending keyframe image request: $imagegen | Each request's output_path exists. |
video_generation | run python: scripts/generate_video.py for each pending shot | clip.mp4 and clip_metadata.yaml exist for each shot. |
rough_edit | run python: scripts/assemble_video.py | final/final_v001.mp4 exists. |
Run loop:
scripts/init_project.py with explicit --id, --title, --target-duration-sec, --language, and brief input.scripts/status.py --project <project> --json.current_stage is complete, stop. If blocked is true, report the blocker.current_stage using the stage table.scripts/validate_stage.py --project <project> --stage <current_stage>; add --concept-id <concept_id> or --shot-id <shot_id> when validating one item. Then rerun status.py --json.Pipeline owns project-level flow and artifact reconciliation. Stage skills own only their scoped outputs.
Bundled Python scripts:
| Script | Use |
|---|---|
scripts/init_project.py | Create a project scaffold. |
scripts/status.py | Read authoritative pipeline state and pending items. |
scripts/validate_stage.py | Run the validator for the current stage. |
scripts/generate_video.py | Generate one shot clip through ComfyUI. |
scripts/start_comfyui.py | Explicitly start ComfyUI from the repo config when requested. |
scripts/assemble_video.py | Assemble generated clips into the rough final. |
Prompt templates:
assets/templates/worker_prompts.md: Use the matching template when creating a subagent for a pending director, concept, shot, or image item.Before every major action, print one line:
[StoryTeller] project=<project_id> stage=<stage> item=<item-or-> action=<action> status=<running|ok|blocked|done>
Use headers only for human-readable progress. Use scripts/status.py --json and validators as the state authority.
Create a subagent for every stage whose executor is create subagent: $skill. Use spawn mode with agent_type="worker", fork_context=false, model="gpt-5.5", and reasoning_effort="medium". Pass only the target project path, target item id, write scope, validator command, and expected output files. Do not fork the full pipeline context into workers.
For concept_requests, concept_images, shot_requests, and keyframe_images, create a fresh subagent for each pending item. If subagents are unavailable, stop and report the blocker. Do not fall back to codex exec or inline skill execution.
Use subagents only for skill stages:
Do not use subagents for status.py, validators, or stages whose executor is run python:.
Each worker task must have one target object and an explicit write scope. Treat files and validators as authoritative, not the worker's final message.
Concept requests and shot requests may run sequentially or in low concurrency when each worker writes only its own target directory.
Stop and report the blocker when:
status.py --json reports blocked: true.validate_stage.py exits nonzero.$imagegen does not create output_path.generate_video.py or assemble_video.py fails.