zenstory drama-skills: Turn a Novel Into an Edited AI Short Drama With 11 Skills

AI short-drama production usually scatters across chat windows, generation sites, and a video editor, with nothing tying the pieces together. drama-skills — a 2k-star, MIT-licensed collection from ZenStory AI — runs the whole chain inside the coding agent you already use, and it comes from production: the README says the team's studio line has run over a thousand AI short-drama and comic-drama projects since 2025.

Why This Skill Matters

Eleven skills cover the pipeline end to end: project init and routing, novel analysis, story development, episode scripts, character and scene assets, image prompts, storyboards with frozen keyframes, video prompts, confirmation-gated production, editing into a finished cut, and review. What holds it together is a file discipline: each episode lives in five Markdown documents — script, visual design, storyboard, image prompts, and video prompts — plus an edit list once you cut. No parallel database; the file you edit is the layer you change.

Character consistency is handled the same way. A visual trait that must hold across shots is written into the visual design document as a "continuity lock" — a phrase that gets pasted verbatim into prompts. A checker script, creator_markdown_check.py, cross-checks the documents and reports causes, not just failures: the README shows it flagging a keyframe that mentions a character the visual notes never covered, and a storyboard duration that disagrees with its video prompt.

The gate that matters commercially: nothing gets generated until you confirm. Every image, video, TTS, or music task is first previewed in the files with exact counts, parameters, and adapters; production runs only after your explicit go-ahead, and vendor credentials stay out of the project.

Installation

One command installs all eleven skills globally (drop -g to scope to the current directory); re-running it updates:

npx skills add zenstory-ai/drama-skills -y -g

The scripts need Python 3.9+ (the version macOS ships is fine), and rendering a finished cut needs ffmpeg/ffprobe on your PATH. Invoke as /short-drama in Claude Code, $short-drama in Codex, or just describe what you want in plain language.

Real Workflow: From a Novel to a Rendered First Episode

Start from a source you own or have rights to adapt. The optional first step is a quick adaptation review: ask the novel-analysis skill for a sampling-based read that tells you whether the book is worth adapting before committing to a full breakdown. Then initialize a project — the README's own example sets up an urban face-slap drama in a vertical 9:16 format.

Write episode one with the script skill, then break it into assets, reference-image prompts, a storyboard with frozen keyframes, and per-shot video prompts. When you name a target video model — MiniMax H3 in the README's examples — the video-prompt skill binds existing character, scene, and prop images as references per shot, or lists exactly which images are missing instead of falling back to text-to-video.

Production is where the pipeline earns its gate: /short-drama-produce previews the exact batch of tasks, waits for your confirmation, then calls external adapters — Seedance, GPT Image 2, MiniMax H3 video, and MiniMax Music are the ones the README names. Finally, /short-drama-edit turns the produced shots into a cut list with in/out points and the reasoning for each, subtitles taken verbatim from the script, loudness normalized, and renders with ffmpeg.

Tips

• A local dashboard comes with the suite: /short-drama dashboard opens a project overview with episode progress and existing media. The README states it listens only on localhost and that project content is not uploaded.
• If a video job's process dies after submission, do not resubmit — tasks are billed at submission time. Run audit, which reports the attempt as orphaned_provider_job, then collect --job-id <id> retrieves the result without paying again.
• MiniMax H3 video prompts print as six sections; the README's FAQ says to copy all of them — they are one prompt submitted once.
• Static comic-drama (keyframe cuts plus voiceover) skips the video-prompt layer entirely; storyboard keyframes and image prompts go straight into production preview.

When Not to Use This

Media generation costs money at the vendor, paid in your own configured credentials — the suite plans and previews, but the billing is yours. The workflow presumes you own or have adaptation rights to the source material. And if you want a one-click text-to-video product rather than an agent-driven production line, an app is the lighter path; this suite trades convenience for control over every layer between script and render.

See the leaderboard for more skills.