A curated bundle of the three most-starred GitHub resources that teach AI agents to draw human anatomy, skeletons, and animals without hallucinating extra fingers, fused limbs, mutated faces, or impossible proportions.
Each pick is vendored into this repo as a git subtree under skills/ so the
full source — prompts, skill files, examples, license, history — travels with
this repo and can be updated cleanly from upstream.
"Bad anatomy" is the single most common failure mode of generative image models: six-fingered hands, two-jointed elbows, mirrored faces, vestigial limbs, animals with the wrong number of legs. Three communities on GitHub have published the most-cited countermeasures:
- A negative-prompt vocabulary that names every anatomical failure mode so the sampler steers away from them.
- An agentic skill (Claude Code / Codex / OpenClaw) that wraps a curated prompt gallery, parameter best-practices, and a CLI — so the model is given structure instead of a blank canvas.
- A prompt cookbook for OpenAI's GPT-Image-2 that pairs every example with the actual rendered output, reference-image anchoring, and explicit "Avoid" clauses so identity and proportions stay intact across edits.
Together they cover the three layers of the stack: what to forbid (negatives), how to ask (skill scaffolding), and what good looks like (reference gallery).
| # | Repo | Stars | Why it made the cut |
|---|---|---|---|
| 1 | EvoLinkAI/awesome-gpt-image-2-API-and-Prompts |
~15.6k | The largest curated gallery of GPT-Image-2 prompts with paired output images, including portrait, character, and figure cases that lean hard on reference-image anchoring + explicit "Avoid: extra fingers, deformed hands, warped face" clauses. |
| 2 | wuyoscar/GPT-Image2-Skill |
~2.4k | The top-ranked agentic skill (Claude Code / Codex / OpenClaw) for image generation. Ships a SKILL.md, CLI, parameter cheatsheets, and a deliberately small, high-signal gallery covering character design, figure illustration, and animal subjects. |
| 3 | mikhail-bot/stable-diffusion-negative-prompts |
~90 | The most-direct match for the brief: a battle-tested vocabulary of negative prompts (bad anatomy, mutated hands, extra fingers, fused fingers, missing limbs, long neck, malformed limbs, disfigured, cloned face) that ships with every modern SD-style stack. Small star count, outsized influence — these tokens are copy-pasted into millions of generations. |
Each lives under skills/<name>/.
Location: skills/awesome-gpt-image-2-API-and-Prompts/
Upstream: https://github.com/EvoLinkAI/awesome-gpt-image-2-API-and-Prompts
License: CC0 1.0
- 583+ curated prompt cases organised by use case (portrait, character design, poster & illustration, e-commerce, ad creative, UI mockups).
- For every case: the exact prompt, the rendered output, and (where relevant) an "Avoid" section listing the anatomical failures the prompt is engineered to suppress.
- API usage patterns for
/v1/images/generationsand the multi-turn editing flow that lets you keep a subject's face/proportions stable across iterations. - 11 language localisations of the README.
| Technique | Example clause |
|---|---|
| Reference-image anchoring | "use the uploaded photo as main reference; preserve the exact facial features" |
| Negative phrasing inside positive prompts | "Avoid: extra fingers, deformed hands, warped face" |
| Identity locking across edits | "Do not alter my facial features, only change the background" |
| Visual over-specification | dense descriptors that leave no anatomical detail for the model to guess |
| Output verification | every prompt is published next to its real generated image, so users can compare expected vs. actual anatomy before reusing the prompt |
⚠️ The authors caution against using outputs as-is in academic figures — always validate anatomy against ground truth.
Browsing for a known visual style and lifting the negative/avoid clauses verbatim into your own pipeline.
Location: skills/GPT-Image2-Skill/
Upstream: https://github.com/wuyoscar/GPT-Image2-Skill
License: MIT
- A complete agent skill folder (
skills/gpt-image/) ready to drop into Claude Code, Codex ($skill-installer), OpenClaw, or Hermes Agent. - A
gpt-imageCLI for batch generation/editing with the OpenAI API. - Markdown reference guides covering parameter semantics (
size,quality,n,format), OpenAI's official prompting best-practices, and a per-domain craft checklist. - A curated (intentionally small) gallery: character design, figure illustration, pixel-art creatures, and photorealistic wildlife.
| Technique | How the skill applies it |
|---|---|
| Curated, not crawled | "Small but mighty — curated for signal, not volume." Every example is hand-vetted, so anatomy patterns aren't averaged across junk prompts. |
| Constraint-based prompts | Each entry uses explicit negations ("no nudity," "no existing copyrighted characters") and bounded descriptors. |
| Local guides bundled with the skill | The agent has on-disk reference markdown to consult before issuing a generation — closes the loop where the model would otherwise improvise parameters. |
| Workflow caution | An on-skill warning instructs the agent to treat outputs as sketches, never as ground-truth anatomy for research figures. |
| Reproducibility | Prompts are versioned alongside seed/parameter metadata so anatomy regressions are catchable in diffs. |
Teams that want an installable, callable skill — not a copy-paste cookbook.
Plays nicely with claude-code plugin marketplaces and Codex skill installers.
# Claude Code
/plugin marketplace add wuyoscar/gpt_image_2_skill
/plugin install gpt-image@wuyoscar-skillsLocation: skills/stable-diffusion-negative-prompts/
Upstream: https://github.com/mikhail-bot/stable-diffusion-negative-prompts
License: see upstream LICENSE
A short, dense list of negative-prompt strings collected from Prodia, Reddit, ImgSli, and community contributors. Every string is engineered around anatomical failure modes. The repo's entire purpose is captured in this single canonical block:
((((ugly)))), (((duplicate))), ((morbid)), ((mutilated)), out of frame, extra fingers, mutated hands, ((poorly drawn hands)), ((poorly drawn face)), (((mutation))), (((deformed))), ((ugly)), blurry, ((bad anatomy)), (((bad proportions))), ((extra limbs)), cloned face, (((disfigured))), out of frame, ugly, extra limbs, (bad anatomy), gross proportions, (malformed limbs), ((missing arms)), ((missing legs)), (((extra arms))), (((extra legs))), mutated hands, (fused fingers), (too many fingers), (((long neck)))
| Failure mode | Tokens it forbids |
|---|---|
| Hands & fingers | extra fingers, fused fingers, too many fingers, missing fingers, extra digit, fewer digits, broken finger, twisted fingers, disfigured hand, poorly drawn hands |
| Limbs | extra limbs, missing arms, missing legs, extra arms, extra legs, malformed limbs, floating limbs, disconnected limbs, missing limb, split limbs, disembodied limb, linked limb |
| Body proportions | bad anatomy, gross proportions, long neck, long body, bad proportions, disfigured butt, extra knee, extra elbow |
| Faces | poorly drawn face, cloned face, 2 heads, 2 faces, cross-eye, squint, deformed face, irregular face, blurred faces |
| General morphology | mutated, mutation, deformed, disfigured, mutilated, morbid, amputee, severed, dismembered |
| Quality / framing | out of frame, cropped, low-res, blurry, jpeg artifacts, watermark, signature |
These tokens generalise: they work for human anatomy, skeletons (the same "extra limbs / fused" failure modes apply), and animals (apply to legs, paws, ears, tails — the same vocabulary catches polydactyl pets and four-eared rabbits).
Dropping straight into the negative prompt field of any Stable
Diffusion, SDXL, Flux, or ComfyUI workflow. Also a useful seed list when
authoring "Avoid:" clauses for closed-source models that don't expose a
dedicated negative-prompt parameter (GPT-Image-2, Nano Banana, Midjourney
v6+ with --no).
┌─────────────────────────────────────────────────────────┐
│ USER INTENT: "Draw a skeleton holding a cat" │
└─────────────────────────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────┐
│ 2. GPT-Image2-Skill │
│ - Agent picks the closest gallery exemplar │
│ - Loads parameter cheatsheet │
│ - Issues a structured prompt │
└─────────────────────────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────┐
│ 1. awesome-gpt-image-2-API-and-Prompts │
│ - Lifts the matching positive prompt + Avoid clause │
│ - Anchors identity with a reference image │
└─────────────────────────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────┐
│ 3. stable-diffusion-negative-prompts │
│ - Appends the anatomy negative vocabulary │
│ - (or: rephrased as "Avoid:" for GPT-Image-2) │
└─────────────────────────────────────────────────────────┘
│
▼
Image (audited against reference)
GitHub repository search on 2026-05-27, sorted by stars, intersected with the following criteria:
- Explicit focus on prompts or agent skills (not model weights, not apps, not 3D viewers).
- Demonstrable coverage of human anatomy / skeletons / animal drawing, either via dedicated cases or via vocabulary that targets anatomical failure modes.
- Active maintenance in the last 12 months (
pushed:>2025-05-27). - Real signal — not Coursera-class fork chains.
Queries that produced the shortlist:
anatomy drawing AI prompt skill in:name,description,readme stars:>10claude skill drawing anatomy svg in:name,description,readmestable-diffusion-negative-prompts sort:starsawesome claude skills sort:stars(for skill-format candidates)
Honourable mentions that lost on either stars or scope:
Sunwood-ai-labs/draw-io-skill— diagrams, not anatomy.Agents365-ai/excalidraw-skill— hand-drawn style, no anatomy gallery.bentossell/visualise— inline SVG, agent-focused, but no anatomy coverage.minimaxir/stable-diffusion-negative-prompt— experimentation notebook rather than a reusable vocabulary.
exec_dashbd/
├── LICENSE
├── README.md ← you are here
└── skills/
├── awesome-gpt-image-2-API-and-Prompts/ (subtree @ EvoLinkAI/main)
├── GPT-Image2-Skill/ (subtree @ wuyoscar/main)
└── stable-diffusion-negative-prompts/ (subtree @ mikhail-bot/main)
Each subtree carries its own LICENSE and README.md; refer to those for
upstream license terms and contribution guidelines.
# Pull latest from any one of the three
git subtree pull --prefix=skills/awesome-gpt-image-2-API-and-Prompts \
https://github.com/EvoLinkAI/awesome-gpt-image-2-API-and-Prompts.git main --squash
git subtree pull --prefix=skills/GPT-Image2-Skill \
https://github.com/wuyoscar/GPT-Image2-Skill.git main --squash
git subtree pull --prefix=skills/stable-diffusion-negative-prompts \
https://github.com/mikhail-bot/stable-diffusion-negative-prompts.git main --squashThis curation/README is released under the repo's existing LICENSE. Each
vendored subtree retains its own upstream license; do not re-license their
contents.