One topic to a finished local episode

Read automation guidance. This skill coordinates storyboard and produce; it does not contain another writing recipe. Use their current contracts. For new episodes follow storyboard's single scenario-v2 → story-v2 flow: causal alternatives, complete acted scenes, blind comprehension, shot planning, then content-derived timing. Read character-led scenarios. For dramatic characters, read character depth before scenario selection. Carry the chosen conflict through performed choices, relationship responses and shot framing; metadata alone does not deliver it. Narrators and incidental roles need no compulsory inner-conflict subplot.

Use legacy unattended notes only to understand a resumed unversioned or story-v1 episode's old files. Never load them as instructions for a new story. No default runtime, shot count, engine mixture, fixed sentence length, mandatory narrator, CTA or one-line-per-shot pattern guides new authoring. Preserve explicit user constraints, real adapter limits and approved budget. Canvas and length are separate choices.

Scope and setup

An invocation authorizes the requested local work; it does not authorize unbounded paid media generation or publication. Reuse existing spending authorization. A storyboard-only or benchmark request stops with plans even if it tests this automatic route.

Read the channel profile and project initialization. Reuse saved genre, model, style, ratio, format, voice and SNS targets. Create/snapshot a new episode with project_manage or its CLI. Existing episodes retain their snapshots and explicit choices. Interactive setup asks missing preferences once. An unattended run with incomplete settings reports project_setup_required and the missing fields, then stops before the lock, authoring or spending (setup report); it invents no choice. Pending SNS connections alone do not block local work. Use model settings and capability_status to check the saved adapters; do not silently switch provider or voice. Generated video goes through video_catalog → video_plan → video_approve (the user's yes) → video_execute under produce's routing contract, never a direct adapter call.

Prepare shared identities and export the character sheets before text authoring. Follow production order: project setup and asset preparation, the story, characters and lines, icon attributes, images, video prompts, then each shot's approved video. Missing authorization blocks generation. story_sequence waits for the episode cast in storyboard/cast.json, each character with its identity and, unless voiceOnly, its face, front and back panels; without authorization to make a missing panel, report it and stop before the story.

For growth-loop calls, read the calling platform's growth-plan.md and follow unattended lifecycle for channel locking, topic selection, duplicate checks, counters and queue scope. Human-invoked local drafting does not inherit a publishing schedule or automatic queue authorization.

Coordinate the steps

Read the production order (user directive 2026-10-31). Only the story order is enforced; checks come back once the quality is known.

1. Storyboard: the story (story_sequence → story_scene → story_shot), each shot's characters and lines (shot_characters → shot_dialogue), then its icon attributes (shot_icons). Prepare props and backgrounds with their images.
2. Invoke generate-cut-images: each shot's start image, and the end image a first_last shot needs, with shot_image.
3. Invoke generate-cut-videos: each video prompt with shot_prompt, then each shot's video after the user's yes (video_approve), or after they approve every shot.

review-context, review-camera, compose-blender, review-cut-images and review-cut-videos run only when the user asks. For a supplied existing episode, regenerate it in this order; preserve its story when the user asks to. Enforce real spending/attempt limits. The coordinator never writes a reviewer's verdict.

After the videos, use produce for assembly, music, requested platform copy and final delivery verification. Keep speech provenance without automatic speech-scoring/STT gates. Publishing remains a separate authorization. Preserve actual costs and failed reports.

Reviews and explicit improvement loops

Reviews run only when the user asks, and none of them gates a step. Do not manufacture scores, freeze stamps or media reviews. Do not invent a passing result to fill a growth queue. Readability, usable files, supported payloads, finite execution timing, concurrency, spending and publication authority remain real requirements.

When the user requests a convergence loop, run that loop to the stated criterion. Keep writers and adversarial reviewers separate and preserve all results. Freeze the rubric before the first read. A request for over 95 requires score > 95 and no P0 for every requested story, not an average across stories. Improve the reusable installed skill and have its authors reload the update and resume saved drafts; restarting an author is optional. Use fresh context for independent evaluation. Work sequentially when requested, carry a resolved lesson into the next story and avoid repeating unchanged checks. Do not quietly repair the evaluation artifacts from the coordinator. A score for a plan describes the plan alone. Without generated video and actual playback, rendered quality is not_generated / not_reviewed regardless of the story score. Evaluation cases cover automatic-route regressions.

Finish and resume

Use produce's episode state and delivery proof for resumes. Do not regenerate an unchanged approved asset merely because a creative score is absent. Track actual spending and failures, release only this run's lock and update queues only within their standing authorization. Present paths, estimated or measured duration, actual costs, review findings and unverified media. For a local-only invocation stop with those artifacts; publishing belongs to the publish skill or the authorized growth loop.