Comfy Prompt Skill

Sister skill to comfyui-cli. This writes the prompt. comfyui-cli runs it.

Language rule: Reply in whatever language the user writes in.

HARD RULES — pre-delivery checklist

Run this checklist BEFORE sending any prompt response. The failure mode it prevents: producing prompts that look correct because of training-data shape rather than verified Comfy model parameters.

Confirm in order:

Routing line first. First line names model + path (cloud/local) + workflow. Example: Routing: cloud → flux-pro (text-to-image, sync, 1024×1024). Missing routing line = response incomplete.
Model exists and is verified. Cloud model must appear in the cloud-models list in frontmatter, OR be a real file in /Users/dawizkidmal/ComfyUI/blueprints/. If user names a model not in either list, say so and offer the closest match. NEVER invent model names. NEVER claim Higgsfield/Runway-app/Pika-app names map without checking the table in model-guide.md.
MCSLA structure intact on video prompts. Model · Camera · Subject · Look · Action. Five layers, every video prompt, unless user explicitly opts out.
Camera/motion vocabulary verified. Camera preset names come from vocab.md. If user names one not in the file, say so and ask for clarification.
Negative constraints appended. Pull from shared/negative-constraints.md. Do not paraphrase — use file phrasing.
Aspect ratio uses correct flag per model family. Comfy Cloud models have different parameter conventions for dimensions:
- BFL Flux (flux-pro, flux-ultra, flux-2): --width INT --height INT
- Stability / Ideogram / Reve / Vidu / Grok: --aspect_ratio STRING
- Seedance: --ratio ENUM --resolution ENUM
- Runway: --ratio ENUM
- DALL-E / Recraft: --size "WxH"
- Pika (t2v): --aspectRatio FLOAT (camelCase!)
- kontext / fill / expand / rmbg / nano-banana: no dimension flag (uses source)
Use the translator script to avoid schema errors:
```
$(python3 ~/.claude/skills/comfy-prompt/scripts/aspect_flags.py <model> <aspect>)
```
Or via cf wrapper: --platform tiktok|reel|wide|square auto-translates. Never hand-write --aspect_ratio — it only works for some models. Anamorphic / 2.35:1 / 2.39:1 are style register (Look line), not output ratios.
Prompt under 200 words. Going over = padding. Tighten.
Execution command appended. End with the exact comfy generate ... or comfy run --workflow ... command the user can paste.

If items 1–8 incomplete, the response is incomplete. Fix before sending.

Two execution paths

Path	When	Skill that runs it
Cloud (`comfy generate <model>`)	Need access to BFL Flux, DALL-E, Seedance, Grok, etc. Fast, no local setup. Requires `COMFY_API_KEY`.	`comfyui-cli` § Comfy Cloud
Local (`comfy run --workflow`)	Have models downloaded, want full workflow control, no per-call cost. Need `comfy launch --background` first.	`comfyui-cli` § Local

Default: cloud for fast prototyping, local for repeat / batch / privacy.

Workflow

Fast Path — Simple Creative Requests

User gives clear creative intent, no specific constraints ("write me a prompt for a car chase at night"). Generate immediately with premium-first defaults:

Parameter	Default
Aspect ratio	16:9
Quality tier	S (premium) — pass `--budget` to downshift to B
Image model (S)	`nano-banana --model gemini-3-pro-image-preview` (Gemini 3 Pro)
Image model (B fallback)	`flux-pro`
Image-text (poster, sign)	`ideogram` (S — strongest text rendering)
Image edit (S)	`flux-kontext-max`
Video model (S)	`kling --model_name kling-v3` (latest, top quality)
Video model (A)	`seedance` (cinematic motion)
Video model (B fallback)	`hailuo`
Image-to-video (S)	`kling-i2v --model_name kling-v3`
Local image	`Text to Image (Flux.2 Dev).json`
Local video	`Text to Video (Wan 2.2).json`
Style	Cinematic
Duration (video)	5s (cloud sync limits), 10s local

Resolve via cf auto <task> "<prompt>" — auto-picks tier-correct model.

Do not ask clarifying questions on Fast Path. Deliver ready-to-paste prompt + command.

Full Path — Production Requests

User signals production intent (multi-shot, specific model, budget concern, client work). Confirm in one message:

Required:

Generation type: Image / Video / Edit / Upscale
Aspect ratio: 16:9 / 9:16 / 1:1 / 4:5 / 3:4
Path: Cloud or Local
Model preference (or ask me to recommend — see model-guide.md)

Optional (skip if user provided):

Visual style
Reference image path (for image-to-image / image-to-video)
Duration (video)
Negative constraints

Ask everything in ONE message — do not split rounds.

Route to the right model — tier system

Full mapping → model-guide.md. Premium-first tier table (S default):

Task	S (premium default)	A	B	C (budget)
`image`	`nano-banana --model gemini-3-pro-image-preview`	`flux-ultra`	`flux-pro`	`nano-banana` (Gemini 2.5 Flash)
`image-edit`	`flux-kontext-max`	`flux-kontext`	`nano-banana`	`recraft-i2i`
`image-text` (poster, sign)	`ideogram`	`nano-banana` (Gemini 3 Pro)	`dalle`	`nano-banana` (Gemini 2.5 Flash)
`illustration`	`recraft`	`ideogram`	`stability-sd3`	`recraft`
`inpaint`	`flux-fill`	`flux-fill`	`recraft-inpaint`	`recraft-inpaint`
`outpaint`	`flux-expand`	—	—	—
`bg-remove`	`recraft-rmbg`	—	—	—
`bg-replace`	`recraft-replace-bg`	`recraft-replace-bg`	`ideogram-bg`	—
`vectorize`	`recraft-vectorize`	—	—	—
`upscale`	`recraft-upscale-creative`	`stability-upscale-creative`	`recraft-upscale`	`stability-upscale-fast`
`video-t2v`	`kling --model_name kling-v3`	`seedance`	`hailuo`	`pika`
`video-i2v`	`kling-i2v --model_name kling-v3`	`runway-i2v`	`vidu-i2v`	`pika-i2v`

Local fallback (for privacy, no-cost iteration, ControlNet):

T2I: Text to Image (Flux.2 Dev).json
I2I edit: Image Edit (Flux.2 Klein 4B).json
T2V: Text to Video (Wan 2.2).json
Upscale: Image Upscale(Z-image-Turbo).json
Inpaint: Image Inpainting (Flux.1 Fill Dev).json

Resolution shortcut: cf auto <task> "<prompt>" [--quality s|a|b|c] [--budget] routes automatically.

Check templates for genre match

Before writing from scratch, check templates/:

User request matches	Template
Chase, pursuit, action, parkour	`templates/01-action.md`
Product, commercial, ad, UGC	`templates/02-product.md`
Portrait, character intro, close-up	`templates/03-portrait.md`
Landscape, nature, establishing shot	`templates/04-landscape.md`
Sci-fi, cyberpunk, VFX, space	`templates/05-scifi.md`
Cinematic still image with shot framing	`templates/06-cinematic-still.md`

Adapt the template — don't paste verbatim.

MCSLA Formula

Five layers, every prompt:

M	C	S	L	A
Model	Camera	Subject	Look	Action

Core rules:

Name specific camera presets from vocab.md
Describe VFX concretely (not "magical effects")
Subject → Action → Camera → Style is the most reliable order
Keep prompts under 200 words

Output format

Single prompt:

**Model**: <model name> (<cloud/local>)
**Aspect ratio**: <ratio>  **Duration**: <Xs>  **Style**: <style>

<Prompt body — MCSLA composed>

**Camera**: <camera preset name from vocab.md>
**Negative**: <pulled from shared/negative-constraints.md>

**Run**:
```bash
<exact comfy generate or comfy run command>


**Two versions (when style varies):**

Version 1 — <Style Name>

<Prompt> **Run**: `<command>`

comfy-prompt