AICG3D 技能加载器

AICG3D_SkillLoader - AICG3D 技能加载器 - is the pack's answer to "my prompts are bad and I don't know why". It ships a library of written creative frameworks - a 3D animation short generator, a brand promo generator, a Chinese celestial palace scene, a co-op game intro, hand-drawn-live-action hybrid, high-density fight choreography, K-pop girl group MV, minimalist product ad, MV subtitles, and an H3 prompt-writing guide among others - and outputs whichever one you pick as text, with your task appended.

The naming will annoy you slightly if you came from agent tooling: these "skills" aren't callable tools, they're long-form prompts. Picking one is equivalent to attaching a director who has already written their brief. That's a fine mental model and it's what the node actually does.

How it composes

The pack merges two source libraries - skills/ (SKILL.md-style packages with reference files) and prompt_guides/ (per-scene prompt plans) - into one catalog, cross-mapping duplicate ids so a single entry can carry both a skill_id and a guide_id. The dropdown labels look like:

H3视频提示词编写 [h3-prompt-writing]
双人游戏开场视频生成器 [co-op-game-intro-generator]
高密度打斗提示词 [high-density-fight-prompt]

Pick one, fill task (or drive it through the optional task_override input from upstream), and the output is:

===== 技能：<name> (<id>) =====
<the skill body>

===== 任务 =====
<your task>

Leave task empty and you get just the header and body - useful when you're feeding it into a chat/LLM node and letting the conversation supply the task.

Outputs are 提示词 (composed - wire this), 技能ID (the stable id, handy for noting which framework produced a render), and 技能正文 (body only).

The "自动选择" default

The default is 自动选择, and it does something specific: it emits an empty body and returns your task text unchanged. The tooltip explains the intent - don't push a framework down, let a downstream LLM pick. That only makes sense if there is a downstream LLM, i.e. you're using the multi-turn dialogue route and the skills list is exposed to it. If you're wiring straight into a segment node's prompt, 自动选择 gives you exactly your raw text and no help at all. Beginners hit this constantly: it looks like the node is broken, and it's doing precisely what it was told.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/JGRFW/comfyui-AICG3D

Restart, hard-refresh, AICG3D category. No model downloads and no Python dependencies for the node itself - it reads markdown off disk. If you actually want an LLM in the loop (for the optimiser or the dialogue path), that's a separate concern: the pack supports an API endpoint or a local GGUF through llama-cpp-python, configured in the prompt-optimiser settings panel, not here.

Where this is genuinely useful, and where it isn't

For structure, it earns its keep. H3 will happily generate something from a one-line prompt and it'll look like it. Give it a framework that specifies single protagonist, single action, one camera move, dialogue pacing and what must not change, and the output becomes controllable. That's exactly the discipline a 5-second clip needs, and having it as a reusable, versionable text blob instead of a saved notepad file is a real workflow gain.

For iteration, it can get in the way. The bodies are long; a segment prompt built from one is not something you can tweak word by word while hunting a specific look. My habit: use the skill to generate a first pass, then copy the good parts into a plain prompt once you know what the shot needs.

And a practical warning that applies to the whole prompt side of this pack: frameworks are written for H3's behaviour - native 24fps, short clips, audio generated with the picture. Hand one to a different video model and you're feeding it instructions it wasn't designed around.

Gotchas

Custom skills need a restart. The dropdown is built from the catalog when the node class registers, so a new folder in skills/ won't appear until ComfyUI restarts.

"Automatic" isn't a downgrade, it's a no-op. Covered above, but it's the #1 confusion with this node.

It doesn't call anything. There's no API key, no network request, no model. If you were hoping this node would write your prompt for you, the prompt-optimiser settings is the thing you're looking for; this node is the shelf, not the librarian.

Skills + optimiser can double up. Running a skill body through the prompt optimiser may compress it. Usually harmless, occasionally it flattens the structure that made the skill worth using. If a shot degrades right after you enabled optimisation, that's the first thing to check.