Creative Skill Router
Purpose
Routes user requests into the correct creative pipeline. The Router does not produce the final prompt; it only does three things: Classify → Load upstream → Hand off downstream.
Request
↓
Task Classification (references/task-classification.md)
↓
personal-identity-profile (always, for any creative task)
↓
aesthetic-director-core (always for creative tasks → produces a Creative Brief)
↓
Specialist (character-design-engine / illustration-direction / image-reverse-analysis / prompt-analysis)
↓
Model Adapter (anima-prompt-compiler / nai5-community-prompt-engineering)
↓
evaluation-loop (on feedback rounds)
Core Rules
- Creative requests must not jump directly to the model adapter. If the user says "Give me an NAI5 prompt with theme X" and X has not been designed yet, first go through
aesthetic-director-corebefore reaching the adapter. Only if the user provides an already completed design (clear character facts, clothing, pose) is it allowed to go directly to the adapter. - Classify by intent, not by keyword. If the user mentions "NAI5", it does not mean the task is "write tags"; it could be "design a character and output it with NAI5".
- Design decisions are separated from model syntax. The adapter does not redesign; the design layer does not write model syntax.
- User's explicit request > Identity profile > Aesthetic direction > Specialist skill > Model syntax.
- Split mixed requests. "Design OC + train LoRA" → complete
character-design-enginefirst, then proceed todataset-management/lora-training. - Feedback rounds go through
evaluation-loop+aesthetic-director-core/references/feedback-diagnosis.md, not direct prompt editing.
Quick Routing Table
| Intent | Trigger Examples | Pipeline |
|---|---|---|
| Character / OC / Clothing design | OC, character design, costume design, standing illustration, haute couture, anime game character | identity → director → character-design-engine → adapter |
| Illustration / Atmosphere / Storytelling | illustration, atmosphere image, vertical screen, half blank space, impressionist style, sense of story, key visual | identity → director → illustration-direction → adapter |
| Existing design → Prompt | prompt, tags, NAI5, NovelAI, Anima (and design is complete) | identity → adapter |
| Reference image reverse engineering | reverse engineer, analyze image, extract prompt, restore style, reference this | identity → image-reverse-analysis → director (if original creation needed) → adapter |
| Prompt review / Optimization | optimize prompt, what's wrong with this prompt | prompt-analysis → (adapter if rewrite needed) |
| Feedback / Iteration | too plain, too messy, doesn't look like it, this version is good, try a different direction | evaluation-loop → feedback-diagnosis → return to failed layer |
| Other image models → Prompt | Midjourney, DALL-E, Imagen, Flux, SD, general, model not specified | identity → (director if unfinished) → general-image-prompt-adapter |
| Technical | ComfyUI, LoRA, dataset, training, labeling | comfyui-workflow / lora-training / dataset-management (status: planned → answer with general knowledge, mark [no module], suggest /new-module) |
| Extension modules | hits 06_extensions/* or trigger words in other module descriptions | That module |
| Non-creative, no module hit | small talk, Q&A, miscellaneous | Do not load any modules; answer directly following kernel §6 tone |
Model selection rules are in references/model-selection.md. The full module index is generated from harness KERNEL §5 (from bundle/manifest.json); docs/skill-registry.md is a human-readable copy of the same data.
Triggers Come from Each Module
The Router does not maintain a hand-written master trigger list. Each module's frontmatter description and metadata.triggers are its own trigger conditions. The build aggregates them into the index. When a new module is added, simply write its triggers properly, and the Router will route to it.
What the Router Outputs
The Router itself does not output long-form content to the user. It decides the pipeline internally and directly starts executing the first step. If classification is ambiguous (e.g., "Help me create a character" without specifying whether a prompt is needed or which model), confirm with one question — do not list a bunch of options.
References
references/task-classification.md— Classification rules and triggersreferences/routing-rules.md— Priority and conflict resolutionreferences/model-selection.md— When to choose Anima / NAI5 / general modelsreferences/execution-flow.md— Standard pipeline and failure recoveryreferences/skill-map.md— Overview of all skill layers