akira-tl/akira-video-skills

Akira Video

akira-video is the Primary Router for Akira Video Skills. It owns the video production entry point, VIDEO.md, current task determination, routing to specialized Skills, and stop boundaries; specific video scripts, reusable materials, shots, generation packages, review, and post-production methods are handled by the corresponding specialized Skills. Full project dependencies and human-agent handoffs follow references/workflow/PRODUCTION-FLOW.md; for existing projects, continue from the actual current work without mechanically re-running upstream steps.

The entire production always prioritizes in this order: user and source facts → story / shot intent (where applicable) → character / object identity and structure → product facts (for advertisements) → spatial and temporal continuity → visual aesthetics → technical flair. Specific levels, conflict escalation, and exceptions for "adopt the actual output" follow references/workflow/AUTHORITY.md; later goals must not violate earlier constraints.

1. Entering or Taking Over a Project

First, read VIDEO.md in the project root. If it does not exist, create the minimum video project structure according to references/project/PROJECT-LAYOUT.md; define the project and current status according to references/project/VIDEO-HOME.md, and follow naming conventions in references/project/NAMING.md. Only create directories and files that are immediately needed; do not pre-generate empty categories. When a generation package is needed for the first time, ensure the target project's .gitignore excludes .tmp/.

When the user provides existing images, videos, audio, rough cuts, or reference renders, first determine their real use case in the current project according to references/project/IMPORT-MEDIA.md; do not force re-running design or generation workflows on directly reusable assets.

Then, based on the current task, read the minimum necessary content:

• Video script and shot overview: video/script/;
• Cross-shot reusable content: video/materials/;
• Current shot: video/shots/<shot-id>/;
• When in post-production for the entire film: video/edit/;
• Upstream novels, chapters, brand materials, or other sources: follow the actual project pointers; do not duplicate the upstream fact source.

Completion criteria: Ability to explain the current video goal, delivery requirements, production content in progress, available reusable assets, current shot / post-production status, and actual blocking issues.

Before proceeding to an expensive production step, run only lightweight gates applicable to the current project according to references/workflow/PRODUCTION-GATES.md; gates must not generate extra project state files. Only when the current project is too large for flat reading, or when creative versions truly need simultaneous comparison, read references/project/SCALING-VERSIONS.md; do not presuppose long-form directories. When cleaning failed results, one-off packages, caches, or old media, verify that the file is not the sole input, currently adopted result, or sole dependency for post-production according to references/project/MEDIA-LIFECYCLE.md.

For how important production decisions are recorded, when VIDEO.md is updated, and why adopted shot results are only written to SHOT.md, follow references/project/RECORDING.md; chat history and one-off generation packages are not long-term project records.

2. Maintaining VIDEO.md

VIDEO.md is the home page of video production meant to be read directly by humans and subsequent agents. It only retains information that is still useful. It consistently covers:

• Project and video goals;
• Delivery requirements;
• Current production progress;
• Current work in progress;
• Blocking issues;
• Key decisions that still affect future steps;
• Navigation to the video script, reusable materials, shots, and post-production.

Ordinary prompt history, all generation results, generation logs, and old decisions superseded by Git history must not be included in VIDEO.md.

See references/project/VIDEO-HOME.md for detailed writing guidelines.

3. Routing Specialized Work

Select the next actual production task based on the user's current goal and existing project content. A complete project does not need to mechanically pass through all capabilities; if a script already exists, you can go directly to shots; if shots and materials exist, you can prepare a generation package; after returning results, you can proceed directly to review.

Load the canonical Skill based on the current task:

• Forming or revising video scripts, character performances, scenes, or shot overviews from novels, chapters, briefs, or other upstream content → video-script;
• Characters or scenes are only functionally described, without specific appearance / costume / material / space / lighting design → video-design;
• Characters need voice consistency across shots, sound references, or the project needs to plan dialogue, ambient sound, sound effects, and overall audio boundaries → video-audio;
• Planning or maintaining reusable character, scene, prop, product prompts and reference materials across shots → video-materials;
• Expanding a shot overview into a SHOT.md, handling action, cinematography, sound, and continuity with adjacent shots → video-shot;
• Writing current generation prompts, selecting a minimal reference set, organizing a one-off generation package, importing user generation results → video-generation;
• After the user returns images or generation results, deciding on adoption, adjusting subsequent steps, post-production fixes, regeneration, or rewriting → video-review;
• Multiple adopted shots entering Premiere Pro, After Effects, DaVinci Resolve, etc., for full-film post-production → video-editing;
• For multi-language subtitles, voiceovers, dubbing, or formal deliverables of the same video → first read references/workflow/LOCALIZATION.md; translation content goes to video-script, audio goes to video-audio, subtitles / graphics / final export goes to video-editing, without duplicating the entire project;
• For brand, product placement, or advertisements requiring verification of product functionality, structure, official materials, and advertising expression boundaries → load the optional video-advertising as needed; if the Skill is not installed, use the explicit coordinate akira-tl/akira-video-skills/video-advertising, hand it to the akira Router / Skiloom to generate a candidate plan and install it after authorization.

The Router only stores the summary of responsibilities above; upon entering any branch, the actually loaded authoritative Skill is the sole source of truth for that professional method.

If the goal requires model-specific adaptation for video models, first read references/workflow/MODEL-ADAPTERS.md to select the minimum adapter package already in the repository; if not installed, hand the explicit coordinate to the akira Router / Skiloom to generate a candidate plan normally and install it. Do not hardcode model parameters or vendor details into akira-video. Image generation currently uses the general rules of video-design + video-materials; no separate image model adapter is maintained.

When external generation has been handed to the user and requires waiting across sessions, record only the generation package path, expected return, next step, and actual blocking issues in VIDEO.md according to references/workflow/WAITING-RESUME.md; on the next turn, prioritize resuming the existing package / returned results, and do not repackage just because the session changed.

4. Human-Agent Boundary for External Generation

By default, users are allowed to perform actual generation in external image/video models. When external generation is needed, the specialized generation Skill creates a one-off generation package from the formal project content:

Formal project content → .tmp/ one-off generation package → user external generation → results return → archive to formal directories → delete temporary package.

One-off generation packages are derivatives and must not become long-term project fact sources; long-term prompts remain in the corresponding material or shot directory, and the temporary package only copies what is truly needed for the current generation.

5. Re-routing and Stop Boundaries

After each specialized Skill completes a bounded action:

1. Archive the returned formal materials, generation results, shot decisions, or full-film project to their corresponding directories;
2. Update VIDEO.md with information that still affects current production;
3. Determine the next production task based on the user's goal and current deliverables.

Stop only when the user's defined scope is complete, the next step must wait for user external generation or material provision, there is a real blocking issue, or the current delivery is complete. When waiting for user external generation, clearly provide the one-off generation package and naming requirements for the return; do not pretend that material has already been produced.