The repository description is two versions and three rules behind
漫剧老李 AIGC 创作全流程 Skill · V6.8 轻量版(Short-Drama Director Suite):面向 AIGC 视听创作的工业化全流程技能包——前置双锁(模型/画幅)、五阶门控剧本、4 View 资产锁定、《资产图册》台账、文武双模分镜、FACS 微表情、七段式投喂提示词、P0~P2 质检与离线分镜看板,38 规则库。适配 OpenClaw·WorkBuddy·豆包Coze·Dify。测试版,二次修改整合开源内容。
At a glance
- What is it?
- A prompt and rule library an AI agent executes for AI video production, whose own summary names an older version and a smaller rule count than the package it describes, whose version history runs to six milestones in twenty days behind a single git tag, and whose slash command list includes an explicit way to switch the gates off.
- Who is it for?
- This is a prompt library, so the useful question is whether its rules are worth obeying. The ones with teeth are the locks that refuse to guess: no model version, no aspect ratio, no writing starts, and a machine check that has to pass before delivery.
- Can I use it commercially?
- Yes. MIT is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository last received commits 14 days ago.
- What is it written in?
- Mainly Python, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on October 3, 2026, and from our analysis. They are not legal advice.
Editorial analysis
The summary text describes an older package than the one shipped
The repository's own one line description names version V6.8 and a 38 rule library. The file in the same repository is titled V6.9.8, the update history's latest entry is V6.9.8 dated 2026-09-20, the structure comment counts 49 files and 41 rule libraries, and the self-check command's own comment counts 41 modules. Four numbers, two of them current and two of them not. The same description is behind on structure as well: it advertises two front locks, for model version and aspect ratio, while the capability table lists three, adding the segment splice method for multi-segment jobs, and it calls the package a test build where the page calls it a lite build. The description also says the work is a second-hand integration of open source content, which is consistent with the older versions having been pushed to a separate archive repository.
Six dated milestones in twenty days, one git tag
The update history is a changelog with dates, running from a multi-agent V6.5 on 2026-09-01 through V6.7 on the 14th, V6.8 on the 15th, V6.9 on the 18th, a combined V6.9.3 to V6.9.7 round on the 19th and V6.9.8 on the 20th. The repository publishes one release, v6.9.8, tagged roughly ten seconds after its last push. So the version numbers here are documentation rather than artefacts, and the difference between two of them is a paragraph of prose rather than something you can install. The V6.9.3 to V6.9.7 entry is the most informative, because it says those five rounds came out of live-shoot retrospectives and names what each one added: a dialogue tone position, a rule that a placed character must be brought into frame, first-shot anchoring with shot size and posture continuity and eyeline assignment, a three-part method for rewriting prohibitions, and a rule forbidding a first shot that puts two subjects in frame as a pair. The history also starts at V6.5, so nothing earlier is described here.
The whole deliverable is one directory
The repository has four top level entries: a gitignore, a licence, this readme, and the skill directory itself. Inside that directory, the structure comment counts 49 files and 41 professional rule libraries, with a SKILL.md at the top acting as the router, the working method and the index of the numbered laws from 0 to 25, and a references folder holding the libraries. Five of those files are named in the listing before it stops. Three carry a star marker: one for the asset-first six stages marked as the process authority, one for the model adapters holding the single authoritative definition of the seven-part format, and one for the aspect-ratio file marked as its own authority, with the H3 adapter file marked for its two checkpoints, field mapping and nine-image sourcing strategy. The distinction between a star-marked authority file and an ordinary one is a governance decision the page does not explain.
One of the fifteen commands exists to skip the confirmation
The command list is where the package's philosophy meets its escape hatch. Fifteen slash commands cover script, asset breakdown, the asset ledger, the character board, the top view, storyboard, dialogue diagnosis, the speech-rate self check, micro-expressions, the emotion curve, video prompt generation, the H3 variant, workflow parameter export, the board and review. The sixteenth entry is not a command for doing work. It is an instruction to skip confirmation and produce a whole episode directly. That sits awkwardly beside the rest of the design, which is built out of refusals: the model version and the aspect ratio are locked before anything is written and defaults are forbidden, a pre-feed checklist runs before prompt handoff, three independent quality gates are numbered P0 to P2, and the delivery rule is that the machine check must come back with zero failures. A package whose main argument is that errors should be stopped before generation ships a command to generate one without stopping.
The prohibition rules exist because negation does not work on video models
This is the most portable idea in the package. The stated observation is that video models are insensitive to negation and sensitive to word stems, so the stem of a banned concept gets rendered even when it is hung off a not or a forbidden. The consequence is a field discipline: the prohibition section carries only negative items, and no positive execution instruction may be mixed in with them, on the reasoning that a positive instruction in that field is an instruction to render the thing. There is a second layer next to it, a safety translation dictionary that rewrites gore and dismemberment into an energy wave scattering, sparks dissipating, or a severe wound freezing on a held frame, explicitly with the note that this lowers platform risk rather than promising a pass. Both are prompt-engineering claims about a specific class of model, which is why the package keeps its model lock at the front of the pipeline rather than treating it as a preference.
Seven prompt parts, and a six part translation for a different model
The feed format is a fixed seven-part structure: frame and style, scene assets, core characters, a blocking declaration, the timeline storyboard, sound effects, and a mandatory prohibitions block. For one model line the package adds the vendor's own bracketing characters for voice, a hit-stop rhythm control, and a choice of two task modes, either full reference or first and last frames. Choosing the other branch converts the same content into a six-part structure with named fields for subject definitions, a summary, retention analysis, a detailed description, an overall soundscape and non-diegetic music, and the conversion is automatic rather than a manual rewrite. The H3 branch also brings its own limits: at most 9 images, 3 videos and 3 audio items, twelve in total, with a documented strategy for picking the 9 images and two separate payloads for handing the work over.
The output check numbers from C1 to C20 on one side and H1 to H12 on the other
Delivery is gated by a script rather than by a reader, and it takes the target model as an argument:
python3 short-drama-director/scripts/validate_prompt.py <稿>.md --model 2.5|2.0|h3On the Seedance side it emits twenty numbered checks, on the H3 side twelve, and the branch is chosen by the model argument, where the syntax shown lists all three accepted values rather than one example. The checks cover duration, dialogue, axis jumps, aspect ratio and asset anchor consistency, plus three levels of relative sound level, and the package adds a rule that after any edit three re-verification steps are mandatory. The count grew visibly in one release: the round driven by live-shoot retrospectives added the dialogue tone position as C16, the placement rule as C17, first-shot anchoring as C18, the prohibition rewrite method as C19 with a companion C19e, and the single-subject first shot as C20, taking the total from fifteen to twenty, alongside a seventeenth rule self-consistency gate.
Consistency is enforced with images, not with descriptions
The first structural decision is that nothing gets storyboarded before it has been drawn. Characters, scenes and props are generated as reference images and locked first, which is what the package calls an asset-first order, and the reasoning is that a description drifts between shots while an image does not. A character gets a four-view board built from ten facial items and six costume and material items, scenes get a set of three space views, and everything is tracked in an asset ledger keyed by character, audio, prop, scene and an open-ended series, graded A, B or C with dependencies that trigger on demand. Two spatial rules keep the storyboard honest: the first shot must carry a named subject, and any character given a position has to be brought into frame in that segment or their exit stated. The earliest version in the history also banned static segments outright, on the grounds that generated video has no capacity for still performance, so emotional build-up has to happen while something is moving.
Editorial conclusion
This is a prompt library, so the useful question is whether its rules are worth obeying. The ones with teeth are the locks that refuse to guess: no model version, no aspect ratio, no writing starts, and a machine check that has to pass before delivery. Read those first, because the summary that describes the package is the least reliable part of it, and because one of the fifteen slash commands is there to bypass exactly the confirmation the rest of the pipeline is built on.
Frequently asked questions
What is the manju-laoli short-drama director skill?
A rule base an AI agent executes rather than a finished screenplay. You drop a SKILL.md and a references directory into an agent's skill folder, and the rules turn script, storyboard and video prompt into eight gated stages, with the model version and aspect ratio locked before any writing starts.
Which video models does the manju-laoli skill target?
Seedance 2.5 and 2.0 as a forced two way choice with their syntax contracts kept separate, plus a format branch for MiniMax H3 that converts the seven part prompt into H3's native six part structure and brings its own numbered checks. One Chinese model is a secondary option, and another was moved out of official support as of V6.8.
How does the manju-laoli skill stop a model rendering something banned?
By rewriting rather than negating. Because models respond to word stems and ignore negation, a banned concept is expressed as its positive opposite, such as an energy wave scattering in place of a body, and the prohibition field is kept to negative items only, with no positive execution instruction mixed in.
What does validate_prompt.py check in the manju-laoli skill?
It is the delivery machine check. It emits twenty numbered checks on the Seedance side and twelve on the H3 side when the model argument selects that branch, covering duration, dialogue, axis jumps, aspect ratio, asset anchors, three levels of relative sound level, and a mandatory re-verification after edits.
Which agent platforms can run the manju-laoli skill?
OpenClaw, WorkBuddy, Doubao agents and Coze, Dify and generic LLM clients, with JSON chunking support to avoid context truncation. It installs with an OpenClaw skills install command pointing at the bundled short-drama-director directory, and ships separate self-check, board and emotion curve scripts alongside the validator.
Official sources
Add this badge to your README
If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.
[](https://hysenlabs.com/projects/lixiaoxiao9888-create-manju-laoli-skill)