漫剧老李 (Manju Laoli) Skill: an AIGC director pipeline for AI agents

The rework problem Manju Laoli is built around

Generating a short drama with a video model fails in a predictable way. A character drifts between shots, the model version changes mid-project and the prompt syntax stops matching, and the error is only visible after a full render. Manju Laoli is a rulebase that tries to move those failures earlier. The README states its three goals directly: no drift, no rework, feed-ready output. The audience is narrow and specific. This is an OpenClaw Skill, or agent skill package, meaning it is consumed by an AI assistant rather than by a person reading a screenplay. The README is explicit that it is not a finished screenplay. You drop SKILL.md plus the references directory into an agent's skill directory and the agent routes requests through the rules. The pipeline is genre-agnostic by design, covering AI short drama, advertising, previsualisation, fight choreography and expression work. What makes it worth a look is the ordering. Asset images are produced and locked before any storyboard exists, so cross-shot consistency has a reference artifact instead of a hope. That ordering is the whole product; everything else is supporting rule text.

How the pipeline actually flows

The README describes a linear chain: lock-in, gated screenwriting, dialogue diagnosis, asset locking, dual-mode storyboard, prompt packaging, independent QC, offline board. Two locks come first and are mandatory. The model version is Seedance 2.5 or 2.0, a forced choice with no default, and the aspect ratio is 16:9, 9:16 or 21:9, also a forced choice. The README states that nothing is written until these are specified. That matters because the prompt syntax diverges between the two model versions, so the lock decides which adapter the rest of the run uses. Screenwriting runs through a five-stage gate: Premise, Structure, Beat, Entity, Page. Assets come next, and the asset system is the most concrete part of the design. A character gets a 4 View asset board covering ten facial features and six clothing material items, and scenes and props get their own images. These are catalogued in an asset ledger with CHR, AUD, PRP, SCN and Uxx identifiers, graded A, B or C, with dependency levels A0 to A3 triggered on demand. Storyboarding then splits into two modes, one for dialogue scenes and one for action, and the action mode uses a 15-second previsualisation score banded R1 realistic, R2 commercial high-energy and R3 fantasy. Output is a seven-section prompt: aspect ratio and style, scene asset, core characters, blocking declaration, timeline storyboard, sound effects, and a mandatory prohibitions block. QC runs as P0 to P2 gates over duration, dialogue, axis crossing, aspect ratio and asset anchor consistency.

Installing the skill and running a first pass

The README gives a quick-start block for OpenClaw and nothing for the other listed platforms. The install command copies the package directory into the skills store under the name short-drama-director. If you are on Coze, Dify or WorkBuddy, the README lists those as supported but does not show an install path, so treat the OpenClaw route as the documented one.

openclaw skills install ./short-drama-director --as short-drama-director

After installing, the README suggests a self-check that runs static and semantic regression checks across the 38 modules. You should expect a pass or fail report rather than a rendered artifact.

python3 short-drama-director/scripts/check_package.py

The minimum execution sequence for the agent is stated in the README: read SKILL.md, complete the A and B locks, set the action intensity band, run the speech-rate self-check if there is dialogue, lock the asset images, storyboard, generate feed prompts against the prompt-feeding checklist, then pass the quality gate review. The commands the agent exposes are slash commands, including /写剧本, /拆资产, /资产图册, /做分镜, /台词诊断, /微表情, /情绪曲线, /生成视频提示词 and /审查. There is also a bypass phrase, 跳过确认，直接出整集, which skips confirmations and emits a whole episode. Two scripts produce artifacts: one compiles a project markdown file into a single-file offline HTML board, and one draws the twelve-beat emotion curve.

python3 short-drama-director/scripts/build_board_lite.py <项目.md>
python3 short-drama-director/scripts/generate_emotion_curve.py

The board script takes the project markdown as its only argument and the README calls that markdown the single source of truth. The README also states that the P2.5 image-generation wording and the P4 feed wording must live on the same board, so the board is a consistency surface, not just a viewer.

Where the rulebase gets in the way

The locks are the first friction point. Model version and aspect ratio have no default, and the README is clear that work does not start until both are chosen. For an exploratory session where you are still deciding what the piece is, that gate is a wall. The second constraint is model support. Kling was moved out of formal support as of V6.8 and is listed as LEGACY, with Jimeng described as a secondary or historical option. If your pipeline is built on Kling, this version is not aimed at you, and the changelog implies you would be working against the grain of the adapters. The third is the nature of the artifact. This is a rulebase for an agent, so its quality is the quality of its instructions, and the README does not publish a benchmark, a sample project or a rendered comparison. There is no way to judge output quality from the repository alone. The compliance dictionary deserves a direct note: the README says it translates gore and dismemberment into phrases like 气浪震散, 火花消散 and 重创定格 to lower platform risk, and then states plainly that it does not promise 100 percent approval. That is an honest boundary, and it also means you cannot treat the dictionary as a guarantee. Finally, the README labels the package a test build assembled partly from second-hand open source material, which is a fair warning about stability rather than a marketing line.

Manju Laoli versus a plain prompt library

The obvious alternative is a folder of prompt templates and a style guide, which is what most people use before they reach for something like this. The difference is state. A prompt library is stateless: you paste a template, you get a prompt, and consistency between shot three and shot forty is your memory's job. Manju Laoli keeps state across the run. The asset ledger, the A0 to A3 dependency levels, the axis lock and the S0 to S4 spatial grading all exist so that a later shot can be checked against an earlier decision. The seven-section prompt format is the other divergence. A template gives you a paragraph; the seven-section format is a fixed schema with a mandatory prohibitions block at the end, which makes the output machine-checkable and makes the P0 to P2 gates possible at all. The cost is rigidity. A template bends to any model; this package is calibrated to Seedance 2.5 and 2.0, and the README says the prompt syntax contract splits cleanly between them, so the format is not portable without adapter work. If you write one-off prompts for a single clip, the schema is overhead. If you are producing an episodic series where a character must look the same in episode four as in episode one, the ledger and the gates are the reason to adopt it.

Maintenance, licensing and what the repository does not say

The last push to the repository was on 2026-09-16, one day before this writing, and the repository is not archived. The changelog shows a fast cadence: V6.5 on 2026-09-01, V6.7 on 2026-09-14, V6.8 on 2026-09-15, with breaking changes in the V6.8 entry, including the Kling demotion, the unification on 4 View character boards and the removal of the trailing parameter line. A cadence that fast with breaking changes means upgrade cost is real: any prompt you generated against a previous version may need rework, and the README does not document a migration path between versions. Licensing is the weaker area. The README badge claims MIT, but the repository root contains only README.md and the short-drama-director directory, with no LICENSE file at the top level, and the package description does not state a licence. The README also says the package is a test build that integrates and modifies open source content. If you plan to redistribute it or ship it inside a commercial product, that combination of an unverified MIT badge, a missing root licence file and reused third-party content is exactly the situation to raise with whoever handles licensing on your side before you build on it. I am not giving legal advice here, only pointing at what the repository does and does not contain.

Editorial conclusion

Adopt it if you already drive Seedance 2.5 or 2.0 through an agent platform and want the asset-lock and QC discipline written down as rules the model has to follow. Skip it if you need a supported product with a licence file, a release history and a stable install path: the README calls this a test build, the repository carries no LICENSE at its root, and the install command is shown only for OpenClaw. Before committing, run python3 short-drama-director/scripts/check_package.py against the checkout and read short-drama-director/SKILL.md to confirm the 38 modules it indexes actually resolve.

Frequently asked questions

Official sources

1. Issues
2. lixiaoxiao9888-create/manju-laoli-skill on GitHub
3. README