image-generation
Skill1 starsChanged 41 days agoMITView on GitHub
AI image generation for any product with OpenAI GPT Image (default) and Gemini, FLUX, Recraft, Ideogram or Seedream when they fit better: illustrations and characters, storybook and app art, marketing and share images, backgrounds and plates, storyboard frames, keyframes for image-to-video, and cutout elements for motion graphics. Writes structured prompts with a locked style suffix, keeps characters, palette and style consistent across 20 to 50 images with cast sheets and references, sizes images for camera moves, makes clean cutouts, reviews its own output like a hostile art director, controls cost, and keeps generated content honest and legally safe. Use it whenever the user wants to generate, edit, restyle, cut out or review AI images, mentions gpt-image, DALL-E, Nano Banana, Midjourney, FLUX or "make an illustration", or asks for consistent characters or a visual style, even if they don't say "image generation". The video-director skill calls it for film assets.
- Reads credentials
62 other projects use a version of this.
Adopted by
Most recently taken up by
- arbrain-ai/arbraintheir own version, first seen Nov 15, 2026
- hoveychen/claw-fleettheir own version, first seen Nov 10, 2026
- danielgwilson/luxintheir own version, first seen Nov 8, 2026
- guinacio/claude-image-gentheir own version, first seen Nov 8, 2026
- CheehooLabs/SpureeSkillstheir own version, first seen Nov 8, 2026
- hex/claude-image-generationtheir own version, first seen Nov 5, 2026
Used with
Projects with it also use gift-evaluator, Podcast Generate and 10 more.
Other files people use alongside this one.
Files found in the same public projects as a version of this one, more often than they turn up anywhere else. Each project counts once, large collections count for less, and this file's own repository is left out.
dav-niu474Skillgift-evaluator
The PRIMARY tool for Spring Festival gift analysis and social interaction generation. Use this skill when users upload photos of gifts (alcohol, tea, supplements, etc.) to inquire about their value, authenticity, or how to respond socially. Integrates visual perception, market valuation, and HTML card generation.
Also in 5 projects that use this
dav-niu474SkillPodcast Generate
Generate podcast episodes from user-provided content or by searching the web for specified topics. If user uploads a text file/article, creates a dual-host dialogue podcast (or single-host upon request). If no content is provided, searches the web for information about the user-specified topic and generates a podcast. Duration scales with content size (3-20 minutes, ~240 chars/min). Uses z-ai-web-dev-sdk for LLM script generation and TTS audio synthesis. Outputs both a podcast script (Markdown) and a complete audio file (WAV).
Also in 5 projects that use this
huang-x-hSkillVideo Generation
Implement AI-powered video generation capabilities using the z-ai-web-dev-sdk. Use this skill when the user needs to generate videos from text prompts or images, create video content programmatically, or build applications that produce video outputs. Supports asynchronous task management with status polling and result retrieval.
Also in 6 projects that use this
huang-x-hSkillVLM
Implement vision-based AI chat capabilities using the z-ai-web-dev-sdk. Use this skill when the user needs to analyze images, describe visual content, or create applications that combine image understanding with conversational AI. Supports image URLs and base64 encoded images for multimodal interactions.
Also in 6 projects that use this
- +1 more from huang-x-h
- +7 more from dav-niu474
Other versions62 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 62
- bytedance/deer-flow★ 85ksame name
- zhayujie/chatgpt-on-wechat★ 48ksame name
- zhayujie/CowAgent★ 48ksame name
- jjyaoao/HelloAgents★ 3.2ksame name
- stophobia/deerflow2.0-enhanced★ 821same name
- langchain-ai/design-agent★ 619same name
- claude-office-skills/skills★ 433same name
- sonichi/sutando★ 396same name
- shinpr/mcp-image★ 191same name
- HybridAIOne/hybridclaw★ 162same name
52 more not listed
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository4 other files it gives its agents.
- scriptwritingSkill
- sound-designSkill
- video-directorSkill
- voice-directionSkill
The file
We split this skill into the rules it gives an agent
2 rules. 0 are common, 2 are distinctive, and 0 are very specific to this project.
0
Common
Most of the 53 edited versions we read keep these.
2
Distinctive
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 53
Generated imagery is never evidence: no real people, no generated product renders passed off as photos, no fake screenshots or events.
- Distinctive: line 99
Every output gets a sidecar (prompt, model, size, references with hashes, usage, request id, cost, style hash).
“Common” is measured: kept by at least 6 in 10 of the 53 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Image generation
Images made for a product are props with a job (an app card, a share image, a plate the camera moves across), not posters.
Model ids churn monthly: check references/models.md dates and the vendor changelog before a new project.
Step 0: read every file before the first action
This skill is loaded whole, never in part. Read this SKILL.md to the end, then every file in the table below, each to
its last line: no offset/limit, no head, grep or skimming, no "only the sections this task needs". If a read comes
back truncated, keep reading until the file ends. The task decides what you apply, never what you read; the table says
where each topic lives, not which files to skip. This holds when the skill is called on its own, from video-director, or
by a sub-agent. If the context is summarised mid-task, read the set again before the next action that depends on it.
| Topic | File |
| Prompt anatomy, the style suffix, edit prompts, negatives per model, style recipes by format | references/prompting.md |
| OpenAI API facts (models, sizes, quality, background, masks, cost), the model chooser | references/models.md |
| Cast sheets, descriptor strings, drift checks, palette locking, style across a set | references/consistency.md |
| Sizes and overscan, light continuity, layers, cutouts, sharpness, image-to-video keyframes | references/motion-assets.md |
| Reviewing your own images, failure table, brief-to-handoff workflow, cost control | references/review.md |
| Ownership, indemnity, people, disclosure, provenance | references/legal.md |
| Script flags, the prompts-file schema, other providers' keys | scripts/README.md |
Ten rules
- Generate at the final aspect, with margin for any camera move, one subject per file when it moves independently, and never with text, logos or UI painted in. Leave surfaces that will carry real text blank ("the card is plain white with nothing on it") and composite the real thing in code.
- Lock a style suffix once (medium, construction, palette names with hex values, light direction, "no text, no logos, no border") and append it verbatim to every prompt. Only the subject block varies. Version it if it must change.
- Cast first. One approved character sheet on a plain background; every scene is an edit that passes the sheet as Image 1 with each character's locked descriptor string repeated word for word. Never route identity through earlier scene images or rejects.
- Describe observable choices (camera height, lens feel, light direction, materials, counts, gaze, who holds what), never "epic, cinematic, stunning, 8k". Mood becomes light and colour; prestige becomes physical detail.
- Disambiguate by physical traits, not names: "a plump hen sitting, small red comb, clearly a hen and not a rooster" fixed a rooster. State gender for gendered nouns; models defaulted "teacher" and "doctor" to women.
- Say what you want first, then name the specific intrusion ("no lettering on the sign"). On models without negative support (FLUX, Gemini) write positives only. A constraint that keeps failing gets stated three ways.
- Cutouts: native alpha only on a model that documents it today (gpt-image-2.5 family with no backdrop words in the prompt), otherwise pure white with "no floor shadow" and matte with BiRefNet (grey or green backdrop when the subject has white in it). Verify alpha in code; review over a saturated colour, never over white.
- Draft cheap (low quality, many variants, a contact sheet), finalise expensive only for picks. Keep every version and a sidecar with the exact prompt, model, size and references.
- Review like a hostile art director: contact sheet first, then zoom on hands, faces, stray text, palette drift, scale between characters, light direction. A pretty image that misses the brief fails. Vision judges hallucinate defects: crop and check the pixels before acting.
- Generated imagery is never evidence: no real people, no generated product renders passed off as photos, no fake screenshots or events. Real product = real screenshots and photos, composited. Never name living artists or trademarked characters. For children's products, check provider terms (the Gemini API terms bar apps likely to be used by minors).
Workflow
- Brief: delivery sizes, shot or asset list with each image's job (plate, cutout, keyframe, card), style references, surfaces that will carry real text.
- Explore: 3 to 4 style directions, each one suffix applied to the same two test subjects, low quality, several variants, one labelled contact sheet; pick with the owner.
- Lock: suffix (versioned), palette, light direction, backdrop line; approve the cast sheet (and a prop or set sheet) at high quality.
- Batch from a prompts file, references first, 2 to 4 requests in flight, never overwriting.
- Review (
review.mdchecklist), then fix: edit when composition and identity are right and one local thing is wrong (one change per edit, restate what to preserve); regenerate when composition, identity or style is wrong, or after two failed edits; make pixel-exact fixes by compositing in code. - Cut out and QA edges at 200% over a saturated colour.
- Hand off trimmed PNGs with a manifest of native size and anchor point; keep prompts next to the images.
Scripts
Paths: scripts live in this skill's scripts/ folder, under the base directory shown when the skill loads (~/.claude/skills/image-generation/ for a standard install; a plugin install puts it elsewhere). The other skills of this suite sit next to it as sibling folders. Paths written as ~/.claude/skills/... below assume the standard install.
uv scripts with --help and --dry-run; keys from OPENAI_API_KEY (other providers in scripts/README.md) or
--env-file. Path: ~/.claude/skills/image-generation/scripts/. Full flags and the prompts-file schema:
scripts/README.md; a product-neutral example prompts file: assets/image_prompts.example.json.
gen_image.py PROMPTS.json [NAMES...] [--out public/img] [--provider openai|gemini|bfl|recraft] [--model ID] [--quality Q]
[--final|--draft] [--aspect 16:9|9:16|1:1|4:5] [--overscan F] [--composite-back] [--regenerate] [--concurrency 2]
[--budget USD] [--batch] [--contact-sheet PNG] [--env-file PATH] [--dry-run]
gen_image.py pick NAME LETTER [--dir public/img]
gen_image.py contact-sheet IMAGES... --out PNG
gen_image.py check-alpha IMAGES...
gen_image.py batch-fetch BATCH_ID [--out public/img]
cutout.py IMAGES... [--out-dir DIR] [--check-dir DIR] [--backdrop auto|#RRGGBB] [--erode PX] [--feather PX] [--margin 12] [--json]- Defaults:
gpt-image-2.5-flarefor drafts,--final=gpt-image-2.5-sunburstat high quality;--draft= low quality, 4 variants at ~1 MP intodrafts/, thenpick NAME bpromotes one.gpt-image-1is refused from 2026-10-23. background: transparentis sent only to the 2.5 family and the output is checked for real alpha (kept but marked FAIL if missing); backdrop words in the prompt or style override the parameter. Otherwise white (orgrey,green,color:#hex,nonefor plates) andcutout.py.- References as
{"ref": "cast", "role": "identity and outfits"}write the "Image N: …" lines for you, identity first;edit_ofiterates on a previous output. Masks apply to Image 1 (white = edit) and are guidance only:--composite-backmakes the edit pixel-exact. - Every output gets a sidecar (prompt, model, size, references with hashes, usage, request id, cost, style hash); the script
warns when a set mixes style hashes. Moderation blocks are never retried; 429/5xx are.
--batchhalves the price for large non-urgent sets. - The Gemini, FLUX and Recraft adapters follow the vendors' documented formats but haven't been run live: check the first real result by hand. Gemini and FLUX take positive prompts only. Recraft supports text-to-image only here.
cutout.pyauto-detects the backdrop, removes the white halo by edge-colour decontamination, warns when the subject shares the backdrop colour or has enclosed holes, skips matting for files that already have alpha, and writes review copies over violet#8342C9.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

