comfy-prompt

Premium-first AI generation workflow for Comfy Cloud. Every task auto-routes to the highest-quality model by default.

A Claude Code skill that turns Comfy Cloud — 50+ image and video models including Gemini 3 Pro, Flux Ultra, Kling v3, Seedance, Ideogram, DALL-E, Recraft, Reve — into a single coherent, tier-aware command-line workflow.

Built by @mvanhorn-inspired prompt structure · v3.2.0 · MIT · Deployed and extended by DaWizKid

Why this exists

Most AI image/video tooling makes one generation easy and ten generations chaos. Each model wants different prompt language. Flux wants cinematic vocabulary and /32 dimensions. DALL-E wants natural sentences and size strings. Kling wants --model_name. Pika wants camelCase floats. Switching models means rewriting prompts. Tracking spend means digging through CLI history. Picking the "right model for the job" means memorizing 50+ schemas.

comfy-prompt fixes that with three structural moves:

Premium by default, budget when needed. Every task auto-routes to the best model. Pass --budget to scale back in one shot.
Tier-aware resolver. S/A/B/C tiers per task type. cf auto image "..." picks Gemini 3 Pro. cf auto video-t2v "..." picks Kling v3. No guessing.
Schema introspection at runtime. New models auto-detected. Dimension constraints applied per-model (Flux requires /32, Pika uses camelCase float, DALL-E uses size strings — handled).

Quick start

# Install
ln -sf ~/.claude/skills/comfy-prompt/bin/cf ~/.local/bin/cf
export COMFY_API_KEY=comfyui-...

# Premium image (auto-routed to Gemini 3 Pro)
cf auto image "matte black coffee mug on wooden counter, golden hour"

# Premium video (auto-routed to Kling v3)
cf auto video-t2v "drone shot over Tokyo at night, neon reflections, 10 seconds"

# Budget mode (downshift everything to B tier)
cf auto image "draft concept exploration" --budget

# Preview tier resolution without spending
cf tiers image           # → nano-banana --model gemini-3-pro-image-preview
cf tiers video-t2v       # → kling --model_name kling-v3
cf tiers image --budget  # → flux-pro

The tier system

Premium quality is the default. Budget mode is opt-in.

Tier	Quality	When to use
S	Best-in-class. No compromise.	Hero shots, finals, client deliverables. Default.
A	Strong premium. ~70% of S quality, ~50% of S cost.	High-volume premium work.
B	Solid mid. Reliable, fast.	Internal review, mood boards, drafts that survive.
C	Bargain.	Iteration, throwaway tests.

Premium S-tier defaults:

Task	Model	Cost
`image`	`nano-banana --model gemini-3-pro-image-preview` (Gemini 3 Pro)	$0.15
`image-text` (poster, sign)	`ideogram`	$0.04
`image-edit`	`flux-kontext-max`	$0.12
`illustration`	`recraft`	$0.03
`inpaint`	`flux-fill`	$0.06
`outpaint`	`flux-expand`	$0.06
`bg-remove`	`recraft-rmbg`	$0.02
`bg-replace`	`recraft-replace-bg`	$0.07
`upscale`	`recraft-upscale-creative`	$0.15
`video-t2v`	`kling --model_name kling-v3`	$0.60
`video-i2v`	`kling-i2v --model_name kling-v3`	$0.60

Full S/A/B/C tier matrix → model-guide.md.

Models supported

Partner	Image models	Video models
BFL (Flux)	flux-2, flux-pro, flux-ultra, flux-kontext, flux-kontext-max, flux-canny, flux-depth, flux-fill, flux-expand	—
Google (Gemini)	nano-banana (gemini-2.5-flash-image, gemini-3-pro-image-preview)	—
OpenAI	dalle, dalle-edit	—
Ideogram	ideogram, ideogram-edit, ideogram-bg, ideogram-reframe, ideogram-remix	—
Stability AI	stability-sd3, stability-ultra, stability-upscale, stability-upscale-fast, stability-upscale-creative	—
xAI	grok, grok-edit	grok-video
Recraft	recraft, recraft-i2i, recraft-inpaint, recraft-rmbg, recraft-replace-bg, recraft-vectorize, recraft-upscale, recraft-upscale-creative	—
Reve	reve, reve-edit	—
Kling	—	kling, kling-i2v, kling-lipsync, kling-extend (v1, v1-5, v1-6, v2-master, v2-1, v2-1-master, v2-5-turbo, v2-6, v3)
ByteDance	—	seedance
Pika	—	pika, pika-i2v
Runway	runway	runway-i2v
MiniMax	—	hailuo
Luma	—	luma, luma-i2v
Vidu	—	vidu, vidu-i2v, vidu-extend
MoonValley	—	moonvalley-t2v, moonvalley-i2v

Plus 60+ local ComfyUI blueprints for Flux, Qwen-Image, Z-Image, NetaYume, Wan 2.2, LTX-2.3 (image, video, edit, inpaint, outpaint, upscale, ControlNet, depth, segmentation, 3D).

Commands

Auto routing (v3.2.0)

cf auto <task> <prompt> [--quality s|a|b|c] [--budget] [--platform P]

Tasks: image · image-edit · image-text · illustration · inpaint · outpaint · upscale · bg-remove · bg-replace · vectorize · video-t2v · video-i2v

Explicit model

cf gen  <model> <prompt> [--tag T] [--platform P] [--async]
cf vid  <model> <prompt> [--image PATH] [--platform P]
cf run  <blueprint.json>            # Local ComfyUI workflow

Recipes (multi-step pipelines)

All accept --quality s|a|b|c and --budget.

cf character "<desc>"                              # 4-angle character sheet with identity locked
cf storyboard "<subj>" "<loc>" "<style>"           # 5-shot cinematic
cf product {lifestyle|3angle|ad} "<product>" ...   # product workflows
cf thumbnail "<topic>" "<scene>"                   # YouTube/social thumbs
cf social <hero.png> "<brand>"                     # 4 social BG variants
cf moodboard "<brief>" [--hq]                      # 9-tile mood reference

Workflow helpers

cf tiers <task> [--quality] [--budget]      # Preview routing
cf lint <prompt> --model M                  # Validate before spending
cf compose <subj> --template T --vertical V --style S    # Merge layers
cf variants <prompt> --axis lighting --exec              # N variations
cf compare <prompt> --models m1 m2 m3                    # A/B test
cf translate <prompt> --to <model>                       # Cross-model adapt
cf embed <asset> --prompt "..." --model M                # Embed metadata in EXIF
cf gallery <dir> [--open]                                # HTML gallery
cf digest week|month                                     # Usage report

Project + monitoring

cf init <path> --name "Project"             # Scaffold project dir
cf refs add <slug> <path>                   # Reference library
cf brand                                    # Show brand.yaml

cf jobs [pending|list|budget|complete|fail]
cf dash [--watch]                           # TUI dashboard
cf watch [--loop]                           # Auto-poll async jobs
cf blueprints list|inspect                  # Local blueprints

Util

cf open [PATH]                              # Open output dir in Finder/explorer
cf models [--raw]                           # List cloud models (enriched / plain)
cf preflight <model>                        # Pre-spend check
cf dedup check|register|list                # Content-hash dedup
cf help

Style variants — same skill, four registers

comfy-prompt includes style snippets in styles/. Same title, four visual languages — all generated via the skill itself.

Ghibli	Film Noir	Cyberpunk	Pixar 3D

See launch-kit/style-variants.sh for the generation script.

Recipes — production multi-step pipelines

Every recipe accepts --quality s|a|b|c, --budget, --dry-run, --retry N, --skip-on-fail.

Recipe	What it does	Est. cost (S tier)
`cf character`	Hero + 3 angles (front, left, right, back), identity locked via Flux Kontext	~$0.50
`cf storyboard`	5-shot cinematic (EWS, MCU, ECU, MS, WS) with shared lighting register	~$0.70
`cf product 3angle`	Hero + 3 angles for product catalog photography	~$0.50
`cf product lifestyle`	Hero + cutout + 4 lifestyle background composites	~$0.50
`cf product ad`	Full Instagram ad pipeline: hero → cutout → BG swap → 4K upscale → animate	~$0.85
`cf social`	4 social-media background variants from a single hero image	~$0.30
`cf thumbnail`	YouTube + Instagram + TikTok thumbnail set (4 aspect ratios)	~$0.32
`cf moodboard`	9-tile mood reference (color, texture, light, subject, environment, composition, mood, scale, motion)	~$0.36-1.35
`cf storyboard`	5-shot cinematic sequence with consistent style	~$0.70

Recipes share recipes/_runlib.sh for retry logic, dry-run mode, tier resolution, and prompt mutation.

Engine flags

Quality & budget control

--quality s              # Premium (default)
--quality a              # High premium
--quality b              # Solid mid
--quality c              # Bargain
--budget                 # Downshift S→B globally for this run

Production safety

--dry-run                # Print pipeline without executing (works without API key)
--retry N                # Retry failed steps up to N times with quality booster mutation
--skip-on-fail           # Continue past failed steps (default: abort)
--platform tiktok|reel|wide|square|cinematic   # Auto-translate to aspect + model-correct dimension flags
--async                  # Submit and return job_id immediately
--tag <tag>              # Custom output tag for organization

Environment

COMFY_API_KEY            # Required. Get one at platform.comfy.org/api-keys.
COMFY_AUTO_OPEN=1        # Auto-open output dir after sync gen
COMFY_OUTPUT_ROOT        # Override ~/Comfy-Output
COMFY_NO_DEDUP=1         # Disable content-hash dedup
COMFY_NO_INTROSPECT=1    # Force hardcoded aspect families (skip runtime schema lookup)

Configuration

Brand config

brand.yaml injects palette / mood / lighting / typography defaults into every prompt via cf compose:

name: "Nova Coffee"
palette:
  primary: "warm-amber-orange"
  secondary: "cream-off-white"
  accent: "deep-espresso-brown"
mood: "cozy artisan ritual"
lighting: "golden hour warmth, soft-focus background bokeh"
typography: "serif headline, sans-serif body"

cf compose "ceramic pour-over kettle" --template product --vertical viral-hook
# → Merges template + vertical + brand.yaml into runnable prompt

Templates + verticals + styles

Layer	Files	Use
Templates (10)	`templates/01-action.md` ... `10-music-performance.md`	Genre shape: action, product, portrait, landscape, sci-fi, cinematic still, horror, fashion, comedy, music
Verticals (10)	`verticals/01-viral-hook.md` ... `10-podcast-visual.md`	Business channel: viral hook, SaaS launch, personal brand, course promo, faceless channel, luxury, before-after, testimonial, AI avatar, podcast visual
Styles (7)	`styles/anamorphic-1970s.md`, `studio-ghibli.md`, `cyberpunk-blade-runner.md`, `film-noir.md`, `pixar-3d.md`, `editorial-vogue.md`, `concept-art-painted.md`	Visual register snippets

How the engine works

User request
    │
    ▼
Step 1 — Task classification (image / image-edit / video / etc.)
    │
    ▼
Step 2 — Tier resolution via scripts/tiers.py
    │  (premium S default, --budget downshifts S→B)
    ▼
Step 3 — Schema introspection (runtime `comfy generate schema`)
    │  (24h cache, catches new models, applies model-specific constraints)
    ▼
Step 4 — Aspect ratio translation (--platform → aspect → model-correct flags)
    │  (6 distinct flag conventions: width/height, ratio enum, size string, camelCase float)
    ▼
Step 5 — Pre-spend lint (catches missing MCSLA layers, off-list models, weak phrasing)
    │
    ▼
Step 6 — Content-hash dedup (identical re-runs return existing output free)
    │
    ▼
Step 7 — Generate via comfy generate (sync or --async with job logging)
    │
    ▼
Step 8 — EXIF metadata embed (prompt, model, params)
    │
    ▼
Output organized into ~/Comfy-Output/YYYY-MM/<tag>/

Tier resolver internals (`scripts/tiers.py`)

from tiers import pick

pick("image")                         # → Pick(model='nano-banana', sub_model='gemini-3-pro-image-preview', ...)
pick("image", quality="b")            # → Pick(model='flux-pro', ...)
pick("video-t2v")                     # → Pick(model='kling', sub_model='kling-v3', sub_flag='--model_name', ...)
pick("video-t2v", budget=True)        # → Pick(model='hailuo', ...)   (downshifted from kling-v3)

Cost discipline

Tier	Image cost	Video cost
S (premium)	$0.04-0.15	$0.40-0.60
A (high)	$0.08-0.12	$0.40-0.50
B (mid)	$0.03-0.08	$0.30-0.45
C (bargain)	$0.01-0.03	$0.30

Cost-saving tactics:

Premium-first defaults — but --budget flag downshifts everything to B tier in one shot.
Content-hash dedup is on by default. Identical re-runs return existing output free.
Pre-spend lint catches missing MCSLA layers and weak phrasing before money leaves the account.
Track spending: cf jobs budget and Comfy Cloud dashboard at platform.comfy.org.
Test locally (free) before committing to cloud video — see local blueprints.

cf jobs budget
# Total spend: $12.61
#
# By model:
#   seedance     12 jobs  $7.20
#   flux-ultra   34 jobs  $3.40
#   nano-banana  201 jobs $2.01
#
# By month:
#   2026-05   247 jobs  $12.61

Architecture

comfy-prompt/
├── SKILL.md                       # Claude Code skill contract
├── README.md                      # This file
├── CHANGELOG.md                   # Version history
├── LICENSE                        # MIT
├── model-guide.md                 # All 50+ cloud models + 60+ local blueprints, tier table
├── vocab.md                       # Camera presets, angles, lighting, motion physics
├── prompt-examples.md             # Worked examples + before/after improvements
├── hybrid-pipelines.md            # Cloud + local stacking patterns
├── brand.yaml                     # Per-project brand config
│
├── bin/
│   └── cf                         # Comfy shorthand wrapper (v3.2.0)
│
├── scripts/
│   ├── tiers.py                   # ★ Premium-first tier resolver (v3.2.0)
│   ├── jobs.py                    # Async cloud job tracker + cost ledger
│   ├── compose.py                 # Template + vertical + brand merger
│   ├── lint.py                    # Pre-spend MCSLA validator
│   ├── preflight.py               # Validate BEFORE spending credits
│   ├── translate.py               # Cross-model prompt translator
│   ├── variants.py                # N-variation generator
│   ├── compare.py                 # A/B test same prompt on N models
│   ├── aspect_flags.py            # Universal model → CLI aspect flag translator
│   ├── schema_introspect.py       # Runtime schema parser + 24h cache
│   ├── schema_cache.py            # Cache layer
│   ├── models_info.py             # Enriched model catalog
│   ├── parameterize.py            # Swap prompt/seed/dims in local workflow JSON
│   ├── organize.py                # Dated output organizer
│   ├── embed.py                   # EXIF metadata embed
│   ├── dedup.py                   # Content-hash dedup
│   ├── gallery.py                 # HTML output gallery
│   ├── dash.py                    # TUI dashboard
│   ├── watch.py                   # Auto-poll daemon for async jobs
│   ├── digest.py                  # Usage report
│   ├── refs.py                    # Reference image library
│   ├── blueprints.py              # Local ComfyUI blueprint browser
│   ├── init_project.py            # Project scaffolder
│   └── __pycache__/
│
├── recipes/                       # Multi-step shell pipelines (8 recipes)
│   ├── _runlib.sh                 # Shared retry / dry-run / TIER helpers
│   ├── character-sheet.sh         # Hero + 3 angles, identity locked
│   ├── storyboard-5shot.sh        # 5-shot cinematic sequence
│   ├── product-3angle.sh          # Product hero + 3 angles
│   ├── product-lifestyle.sh       # Product × 4 lifestyle scenes
│   ├── instagram-ad.sh            # Full IG ad pipeline (hero → cutout → BG → 4K → animate)
│   ├── social-pack.sh             # 4 social BG variants
│   ├── thumbnail-set.sh           # YouTube + IG + TikTok thumbnail set
│   └── mood-board.sh              # 9-tile mood reference
│
├── templates/                     # 10 genre templates
├── verticals/                     # 10 business-channel templates
├── styles/                        # 7 style register snippets
├── shared/
│   └── negative-constraints.md    # Negative-prompt phrases by category
│
├── launch-kit/                    # Demo + marketing assets
│   ├── poster.png                 # Wide hero (Higgsfield Z Image)
│   ├── demo-reel.mp4              # 10s generative gallery (seedance)
│   ├── poster-{ghibli,noir,cyberpunk,pixar}.png   # Style variant pack
│   ├── style-variants.sh          # Generation script for variants
│   ├── hero-promo.sh              # Hero generation script
│   ├── demo-video.sh              # Demo reel script
│   ├── linkedin-post.md           # Launch post copy
│   ├── x-thread.md                # X thread copy
│   ├── one-liner.md               # 1-line pitches
│   ├── quickstart.md              # Onboarding guide
│   ├── fork-and-customize.md      # Fork workflow
│   └── press-kit/                 # Brand assets
│
├── tests/                         # 38 pytest + 9 bats integration tests
│
└── mcp_server/                    # MCP integration (optional)

Requirements

Python 3.12+
Comfy Cloud CLI: pip install comfy-cli (or follow platform.comfy.org install)
Node.js (for some local ComfyUI blueprints)
Claude Code with skill loading enabled

python3 --version    # needs 3.12+
comfy --version

Optional:

ImageMagick (brew install imagemagick) for recipe grid composites
jq for JSON parsing in some scripts

Testing

cd ~/.claude/skills/comfy-prompt
python3 -m pytest tests/ -v       # 38 unit tests
bats tests/test_recipes.bats      # 9 integration tests (mocked, zero spend)

All recipes support --dry-run for offline pipeline verification without API key.

Examples

Premium image with Gemini 3 Pro

cf auto image "matte black coffee mug on dark wood counter, golden hour, shallow depth of field, photoreal cinematic" --platform square
# → nano-banana --model gemini-3-pro-image-preview
# Output: ~/Comfy-Output/2026-05/image/nano-banana_<timestamp>.png

Cinematic video with Kling v3

cf auto video-t2v "drone shot pulling away over rain-soaked Tokyo alley, neon signs reflecting on wet pavement, atmospheric haze, 10 seconds, anamorphic cyberpunk" --platform wide
# → kling --model_name kling-v3
# Async job logged. Resume: cf watch --loop

Full Instagram product ad pipeline

cf product ad "matte black coffee mug" "sunlit Scandinavian kitchen, plants in window"
# Step 1/5: Hero on white seamless (nano-banana --model gemini-3-pro-image-preview) — premium
# Step 2/5: Background remove (recraft-rmbg)
# Step 3/5: Lifestyle background (recraft-replace-bg)
# Step 4/5: Upscale to 4K (recraft-upscale-creative)
# Step 5/5: Animate vertical clip (kling-i2v --model_name kling-v3, async)
# Total: ~$0.92

Budget mode — same pipeline at 1/3 the cost

cf product ad "matte black coffee mug" "sunlit Scandinavian kitchen" --budget
# Same pipeline, B-tier models throughout.
# Total: ~$0.31

Cross-model prompt translation

cf translate "cinematic 70mm anamorphic widescreen of rain-soaked Tokyo alley" --to dalle
# → Re-phrased for DALL-E natural-sentence style with size hint

A/B test across models

cf compare "futuristic Tokyo at night, neon reflections on wet pavement" \
  --models "nano-banana:gemini-3-pro-image-preview" flux-ultra ideogram \
  --aspect_ratio 16:9
# Side-by-side output for visual comparison

License

MIT — see LICENSE

Credits

Built and maintained by: DaWizKid — former music producer turned AI developer, Atlanta metro. Building AI-native workflows on Claude Code + Comfy Cloud + Higgsfield.

Prompt structure inspired by: Higgsfield AI prompt skill — original MCSLA framework.

Related skills

Skill	Description
last30days	Multi-source social media research engine — Reddit, X, YouTube, TikTok, HN, GitHub, Bluesky
`comfyui-cli`	Sister skill that handles raw `comfy generate` and `comfy run` execution
`prompt-master`	Optimized prompt generation for any AI tool

Built for Claude Code. Works with OpenClaw, Hermes Agent, Codex, and any agentic runtime that supports skill invocation.

Name		Name	Last commit message	Last commit date
Latest commit History 8 Commits
assets		assets
bin		bin
launch-kit		launch-kit
mcp_server		mcp_server
recipes		recipes
scripts		scripts
shared		shared
styles		styles
templates		templates
tests		tests
verticals		verticals
.gitignore		.gitignore
CHANGELOG.md		CHANGELOG.md
LICENSE		LICENSE
README.md		README.md
SKILL.md		SKILL.md
brand.yaml		brand.yaml
hybrid-pipelines.md		hybrid-pipelines.md
model-guide.md		model-guide.md
prompt-examples.md		prompt-examples.md
vocab.md		vocab.md

Folders and files

Latest commit

History

Repository files navigation

comfy-prompt

Why this exists

Quick start

The tier system

Models supported

Commands

Auto routing (v3.2.0)

Explicit model

Recipes (multi-step pipelines)

Workflow helpers

Project + monitoring

Util

Style variants — same skill, four registers

Recipes — production multi-step pipelines

Engine flags

Quality & budget control

Production safety

Environment

Configuration

Brand config

Templates + verticals + styles

How the engine works

Tier resolver internals (scripts/tiers.py)

Cost discipline

Architecture

Requirements

Testing

Examples

Premium image with Gemini 3 Pro

Cinematic video with Kling v3

Full Instagram product ad pipeline

Budget mode — same pipeline at 1/3 the cost

Cross-model prompt translation

A/B test across models

License

Credits

Related skills

About

Topics

Resources

License

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Uh oh!

Contributors

Uh oh!

Languages

Tier resolver internals (`scripts/tiers.py`)

Packages