Stop AI Hand-Drawn Style Drift: Two Open-Source Solutions — Style Supermarket vs. Precision Recipes

The biggest pitfall in AI hand-drawn generation is style drift — the same style works once, but change the subject and colors, lines, and mood fall apart. To reproduce a style reliably, you need an extremely specific, repeatable prompt every time. This article reviews two open-source projects that turn hand-drawn styles into recipes so you supply the content and they supply the locked-in prompt.

Quick comparison: supermarket vs. precision workshop

handraw-style (by yang0, 4,473 stars) takes an exhaustive approach: 326 hand-drawn styles, 161 layout templates, and 36 classic monochrome palettes, all numbered. Pick a style number, a layout number, and a color number, and it emits a bilingual prompt ready for any image model.

hand-drawn-styles (by liulei, 1,423 stars, MIT) takes a curated approach: only 19 recipes (18 base + one validated variant 3.1) that have been repeatedly verified. Its value isn't quantity but engineering rigor — each recipe is a reproducible artifact with anchor-image validation and a three-stage correction workflow.

One is a massive style supermarket where you browse and combine; the other is a custom workshop where every item comes with tolerances and acceptance criteria.

handraw-style: pick a number, get a prompt

The project's one-liner: can't describe a style? can't compose a layout? pick a number, get a bilingual prompt instantly.

Three combinable layers

Style layer: 326 illustration styles (001–326) grouped A–H: A international editorial cartoons, B international picture-book narrative, C modern graphic/artistic figures, D Japanese contemporary, E Chinese contemporary, F general web/regional, G Chinese contemporary supplement, H other curated. Each style has a numbered mosaic reference so you don't guess.

Layout layer: 161 templates in five categories: social-media cards (21), data infographics (35), comic panels (68), IP design (13), e-commerce layouts (24).

Color layer: 36 classic monochromes (C-01–C-36) across six families: classic blue, fresh green, classic red-green, romantic pink-purple, warm earth, neutrals — including Klein Blue, Sage Green, Hermès Orange, Payne's Grey.

Real example: Autumn Equinox poster

Layout: Poster Design Style: 268 (Contemporary Humanistic Ink Comic) Main Color: C-26 Persimmon Orange + Accent C-03 Prussian Blue Theme: Autumn Equinox — day and night equal, cold and heat balanced

The generated prompt includes scene, audience, content density, emotional tone, main color, editorial style, and style traits — drop it into an image model. The article shows the actual output: large calligraphic title, poetic couplets, seals, day/night split imagery, and relaxed ink figures fused into one composition where text and image are integrated, not pasted.

Image-text vs. pure-image mode

A toggle lets you choose: pure-image mode lets the subject drive the visual content (for illustrations without text); image-text mode preserves your written theme so text participates in composition (for quote cards, social media). Same numbers and theme, switch modes without re-picking styles.

Extended workflows

Long-form illustration: feed the full article; it picks 2–5 visual turning points (opening thesis, abstract metaphor, concept breakdown, process walkthrough), locks one style and palette for the whole piece, batch-generates, and auto-inserts images back into Markdown.

Cover design: give full text or video subtitles; it extracts a 200-word summary and audience persona, assigns style and palette, injects instructions like "design metaphor first, main title prominent, minimal small text."

Photography planning: 10+ shoot types (couple, wedding, portrait, family). First round delivers three distinct region/worldview concepts, each wrapped in an independent code block ending with "output 8-panel grid"; reply A/B/C to proceed.

Installation & natural-language use

Help me install this Skill: https://github.com/yang0/handraw-style

Don't want to browse the catalog? Just say: "Generate a healing illustration prompt of a cat sunbathing on a windowsill, you pick style and palette." It matches a combo and explains why, e.g., "Style 018 · Minimal Deadpan Dialogue Cartoon + Palette C-26 Persimmon Orange (reason: warm, healing, relaxed feel)."

Model-adaptation strategy

The tool only produces prompts; generation strategy depends on the target model. For deeply adapted models like gpt-image, it prefers "author name + style name + core traits" invocation, falling back to reference images only if needed. For non-calibrated models (Midjourney, Flux), it goes straight to reference-image fallback to guarantee fidelity.

hand-drawn-styles: lock style into an engineering artifact

If handraw-style is a style dictionary, hand-drawn-styles is a reproducible recipe protocol .

Tool-agnostic core

Only two plain-text files: PROTOCOL.md (execution flow) and STYLES.md (style recipes). No platform lock-in — works with Claude Code, Cursor, Codex CLI, Gemini CLI, Cline, Windsurf. Their SKILL.md and AGENTS.md are thin adapter shells pointing to the same two core files.

Single responsibility: style only

It explicitly ignores writing, publishing, etc. You give content requirements; it returns a clean, copy-pasteable final prompt. Image generation is delegated to gpt-image, Jimeng, or Midjourney; it does not generate images itself.

19 recipes in eight groups

Realistic hand-drawn: 3 Ghibli, 3.1 Crayon Childish Scrawl (messy self-portrait variant), 7 Crayon Photo-Real, 10 Emotional Narrative Light Wash Sketch

Line·Explanation·Sketch: 1 Minimal B&W Line (xkcd stick figures), 4 Little Bean Doodle Infographic, 6 Ballpoint Single-Line Doodle

Deliberately Bad: 2 Crayon Childish, 5 MS Paint Bad Doodle

Traditional·Retro Texture: 8 Ink Wash Expressive, 9 Retro Pixel

Animation·Concept Design: 11 2D Watercolor, 12 Warm Light Children's Illustration, 16 Spotlight Gouache Character, 17 Ink Line Picture Book

Paper Craft·3D Handmade: 13 Nordic Paper Cut

Picture Book·Flat & Nordic: 14 Nordic Picture Book Gouache, 18 Warm Flat Picture Book

3D·Designer Toy: 15 Big Nose Soft Doll

Invocation

Say "Draw a cat holding an umbrella in rainy day with Ghibli style" or "Use style 3 to draw…" or "Use ghibli to draw…" — Chinese/English aliases and numbers all work. If you specify an aspect ratio it uses yours; otherwise it doesn't force one.

Engineering-grade stability: the 3.1 recipe deep dive

Recipe 3.1 "Crayon Childish Scrawl — Messy Self-Portrait" requires the anchor image anchor-family.png to be sent as a pure style reference with every generation request. Only line quality, facial features, body proportions, crayon stroke trajectories, and negative space may be referenced; character identity, clothing, pose, and narrative must not be copied — style inherits, content is brand new.

Three-stage workflow

After base generation, two mandatory correction passes that only modify crayon strokes : 1. Gentle pass to break up uniform texture. 2. Targeted pass for residual regular hatching, enlarged blunt strokes, and large blank areas. Only the third stage output is the final art. The project explains why: a single strong correction often arbitrarily changes clothing colors or character shapes, so a single-pass substitute is forbidden.

Anchor image verification

Even the anchor image itself is validated: fixed dimensions, decoded pixel SHA-256 must match; metadata changes don't count as style identity change. This level of constraint makes it more like a manufacturing spec with acceptance tests than a prompt trick.

Python renderer for exact extraction

python3 scripts/render_prompt.py \
  --style 3.1 \
  --subject 'Dad puts snack bag back in cabinet, boy stands watching' \
  --text 'No text at all' \
  --aspect 3:4 \
  --format json

The renderer outputs a formal JSON call package containing the prompt, input replay, and required references, preventing agents from abbreviating or mixing recipes.

Use both — they don't conflict

handraw-style solves "too many choices." 326 styles, high-frequency updates (still active early October), broad coverage: Xiaohongshu covers, official account headers, long-form explainers, multi-panel comics, e-commerce main images, photography planning. Official Feishu knowledge base adds quick-reference gallery, Q&A, and commercialization cases; community leans operational/commercial.

hand-drawn-styles solves "too hard to reproduce." Only 19 battle-tested recipes. For when you don't need a hundred options but need one style to come out identical every time, integrated into your existing toolchain, verifiable, pipeline-ready. Fixed recipes, verbatim renderer extraction, hash-verified anchor images — all three designs aim to turn style from "mysticism" into "industrial standard."

Selection guide

Want breadth and immediate firepower — many styles and layouts, quick output → handraw-style.

Want precision and reproducibility — lock a few styles into production pipeline → hand-drawn-styles.

They complement each other: explore rapidly with handraw-style to find direction, then lock the chosen style with hand-drawn-styles' protocol for batch production.

Installation

Both use "send repo URL to your agent" installation.

handraw-style

Help me install this Skill: https://github.com/yang0/handraw-style

hand-drawn-styles

Help me install this Skill: https://github.com/threerocks/hand-drawn-styles

Manual integration per tool: Claude Code — drop entire repo into ~/.claude/skills/hand-drawn (SKILL.md auto-triggers); Cursor — place two core files in .cursor/rules/; Codex CLI, Gemini CLI — point AGENTS.md; Cline, Windsurf — merge into their respective rules or custom instructions. After integration, say "Draw in hand-drawn style…" to trigger.

Community contribution

handraw-style accepts new styles via scripts ( import_tweet_style.py, import_manual_style.py). hand-drawn-styles sets a higher bar: to add a style you must update STYLES.md with the recipe, register menu entry and aliases in PROTOCOL.md, attach 1–2 output samples in examples, and verify at least two cases — "reference-image isomorphic scene" plus "one cross-subject scene" — before merge. High barrier ensures every item on the shelf is genuinely validated.

One pushes breadth to the limit; the other drives a single point to rock-solid. For AI hand-drawn generation, these two repos represent two equally valid philosophies.