One Open Source Project a Day (No. 230): HandRaw-Style — Numbering 327 Hand-Drawn Styles, 165 Layouts, and 36 Color Palettes So AI Art Stops Drifting Every Time

Introduction

"It's not that you can't describe a style well enough — it's that style was never something you should describe on the fly in the first place."

This is the 230th article in the "One Open Source Project a Day" series. Today's project is HandRaw-Style.

One of the most frustrating experiences with AI image generation goes like this: you finally stack enough adjectives together to get a satisfying "healing-style hand-drawn illustration," then switch the topic and regenerate — and the whole style drifts. Line weight changes, the color tone shifts, even that "healing" feeling disappears. The problem isn't that your prompt wasn't detailed enough — it's that describing a style in natural language is inherently ambiguous. Feed the same sentence to a model twice, and the interpretation can still land in different places.

HandRaw-Style's approach is direct: if the ambiguity comes from describing things on the fly, then stop describing on the fly. Pre-catalog 327 hand-drawn illustration styles, 165 layout templates, and 36 classic color palettes, all with tested bilingual prompts attached to each number. Want a particular style going forward? Just say the number — no more improvising adjectives, and the style naturally stays consistent.

4.7k Stars, distributed as a Skill, installable and callable directly by AI coding assistants like Codex and Claude Code.

What You Will Learn

• How HandRaw-Style replaces vague natural-language style descriptions with a numbering system
• Three independent numbering systems: styles (001–327), layouts (SC/IC/CS/IP/EC prefixes), colors (C-01 to C-36)
• The difference between smart recommendation mode and precise selection mode
• Specialized working modes for poster design, article illustration, and photography planning
• The model capability matrix: differentiated adaptation strategies across image models

Prerequisites

• Experience using at least one AI image generation tool (Midjourney, DALL-E, Stable Diffusion, etc.)
• Familiarity with the Skill/plugin mechanism of AI coding agents (e.g., Claude Code, Codex)
• Optional: having experienced "endlessly tweaking a prompt but the style still won't stay consistent" firsthand will make this hit closer to home

Project Background

What It Is

HandRaw-Style's official positioning is "a numbered gallery of hand-drawn art styles paired with a bilingual AI-prompt-generation skill." It's not a traditional code project in the usual sense — it's a structured prompt asset library packaged as a Skill that AI agents can install and invoke directly. Its core value proposition is solving the "vague style description, every generation drifts" pain point, by replacing improvised, on-the-fly descriptions with pre-numbered, pre-tested style/layout/color combinations.

Team and Background

• Author: yang0
• License: a permissive custom license requiring attribution (free to use, modify, and commercialize, with a request to retain attribution to author yang0 and the original repo URL)
• Distribution format: an AI Skill package (compatible with Codex, Claude Code, and similar coding assistants), paired with an offline HTML gallery

Project Stats

• ⭐ GitHub Stars: 4,700+
• 🍴 Forks: 549
• 👀 Watchers: 14
• 📄 License: Custom permissive license (with attribution requirement)

What It Does

The Problem It Solves

The traditional way of describing art style for AI image generation:
  Users improvise a natural-language style description every single time
  ↓ "healing-style hand-drawn illustration, clean lines, warm tones..."
  ↓ the model's interpretation of the same description carries inherent randomness
  ↑ switch the topic once, and the style has already drifted
  ↑ accumulated "prompting intuition" can't be reused — you re-explore from
    scratch next time
 
HandRaw-Style's approach:
  Pre-test and catalog 327 styles + 165 layouts + 36 color palettes, all numbered
  ↓ users just state a number + a topic
  ↓ the output is an already-validated bilingual prompt template
  ↑ style/layout/color stay stable and reproducible, no longer dependent on
    on-the-fly description
  ↑ "prompting intuition" becomes a queryable, reusable numbered asset

Use Cases

1. Social media cover design

   • For Xiaohongshu or WeChat cover images, specify a style code + color code directly to quickly produce platform-appropriate visuals

2. Illustrations for long-form articles / knowledge content

   • The batch illustration auto-fill mode plans 2–5 illustration insertion points, locks one consistent style across the whole piece, and auto-formats into the final Markdown file

3. Infographics / data visualization

   • The "universal poster thinking method" treats infographics as high-density posters, supporting custom timelines and data displays

4. Comic storyboards and IP design

   • Among the 165 layout templates are dedicated comic storyboard (68 variants) and IP design (13 variants) templates

5. Photography / portrait shoot planning

   • The professional photography planning mode generates 3 candidate concepts and an 8-grid layout for couples, weddings, and portrait sessions, requiring a mix of real-world landmark concepts and a fantastical/time-travel concept

Quick Start

Installation (conversational, no command line needed):

Just send this to Codex or any AI agent that supports Skills:
 
"Help me install this Skill: https://github.com/yang0/handraw-style,
 then use it to help me design illustrations."

The agent automatically pulls the repo and configures the style/layout/color resources.

Usage examples:

# Precise selection mode
Style #041, topic: autumn's first milk tea
 
# Combine layout + style + color
Layout: SC-001, style: 041, color theme: C-01, topic: autumn's first milk tea
 
# Poster design mode
Please give me a poster prompt, topic: the autumn equinox

Core Features

1. Three Independent Numbering Systems

2. Two Rendering Modes

• Pure-image mode: text-free illustration output
• Text-integrated mode: text woven directly into the composition

3. Five Specialized Working Modes

4. Model Capability Matrix

A differentiated adaptation strategy across image generation models: for explicitly deep-tuned models (e.g., gpt-image-2), it outputs finely-tuned prompts directly; for general-purpose models like Midjourney, Flux, Stable Diffusion, Imagen, and Gemini, it falls back to a "reference-image high-fidelity fallback" strategy, using a 4-grid reference image to help align the style. This capability matrix is stored at skills/handdraw-style-prompter/references/model_capabilities.json, and is open to community-contributed test data.

A Deeper Look

Turning "Style Description" From a Language Problem Into a Retrieval Problem

HandRaw-Style's core design insight is really a reframing of the problem itself: instead of trying to make natural-language description more precise (e.g., writing longer, more detailed prompt templates), it bypasses the description step entirely, converting "picking a style" from an open-ended language-generation problem into a closed-set numbered-retrieval problem.

Open-ended problem (the traditional way):
  "I want something... maybe healing-style, but not too sweet, with
  simpler lines..."
  ↑ the description space is effectively infinite — the same intent
    can be phrased in countless ways
  ↑ the model's interpretation can land in a different spot in that
    space every time
 
Closed-set problem (the HandRaw-Style way):
  "Style #041"
  ↑ corresponds to an already-validated, fixed prompt template
  ↑ the only remaining randomness comes from the model's own generation
    variance — not from "interpreting the description"

The value of this reframing: it eliminates half the sources of instability. AI image generation inherently has generation-level randomness, which can't be removed at the model layer. But "interpretation randomness" from the description step can be removed by pre-fixing the description. That's exactly what HandRaw-Style does.

Why the Three Numbering Systems Are Kept Separate

Style, layout, and color are designed as three independent numbering dimensions rather than merged into one giant lookup table — this design choice is worth examining:

If merged into one giant table:
  Combinations of "style041 + layout SC001 + color C01" =
  327 × 165 × 36 ≈ 1.94 million combinations
  ↑ testing a prompt for every single combination is simply not feasible
 
Kept as three separate dimensions:
  Styles, layouts, and colors are each independently tested and validated
  ↑ users freely combine as needed — the three dimensions are orthogonal
  ↑ maintenance cost is 327+165+36=528 tested entries, not 1.94 million

This is a textbook example of "using orthogonal decomposition to shrink the state space" — style is about line work/brushstroke/emotional tone, layout is about composition/information density, color is about chromatic relationships. These are relatively independent variables in visual design to begin with, so numbering and validating them separately both lowers maintenance cost and leaves the user's freedom to combine them uncompressed.

The Design-Process-Front-Loading Behind "Design the Metaphor First, Then Generate"

The article/video cover design mode has a rule worth unpacking: "design the metaphor first, then generate; prominent main title, minimal or no subtext." Behind this sentence is actually a front-loaded professional design workflow:

The amateur approach:
  Stuff the article title directly into the image as text
  → crowded layout, lacking any sense of design
 
HandRaw-Style's built-in approach:
  Step 1: distill a visual metaphor from the ~200-character summary
          (not the summary itself)
  Step 2: compose the image around that metaphor, keeping the main
          title clean and prominent
  Step 3: reduce or remove subtext to avoid information overload

This effectively hard-codes a seasoned designer's habit of "think of the metaphor first, worry about layout second" into the Skill's default execution order — users don't need to understand design methodology themselves; they just invoke this mode, and the process has already front-loaded that professional judgment for them.

The Choice of Distribution Format: a Skill, Not a SaaS or Plugin

HandRaw-Style chose to distribute as a "Skill package," riding on existing AI agent infrastructure like Codex and Claude Code, rather than building its own website or image-generation frontend. The benefit of this choice is direct: it doesn't need to integrate with image-generation APIs itself, doesn't need to maintain servers, and doesn't need to handle authentication or billing — all the actual image generation is still handled by the AI tools and image models users already use. HandRaw-Style's only job is providing "numbered, pre-tested prompt assets." It's a fairly disciplined scope boundary: don't overreach into image generation itself, just do prompt asset management well.

How It Compares to Generic Prompt Template Libraries

Project Links and Resources

Official Resources

• 🌟 GitHub: https://github.com/yang0/handraw-style
• 📄 License: Custom permissive open-source license (with attribution requirement)
• 📚 Docs: STYLES.md, LAYOUTS.md, COLORS.md, TUTORIALS.md within the repo

Related Resources

• skills/handdraw-style-prompter/gallery/ — the offline interactive gallery (index.html / layouts.html / colors.html / tutorials.html)
• skills/handdraw-style-prompter/references/model_capabilities.json — the model capability matrix, open to community-contributed test data

Summary

Key Takeaways

1. Turns style description from a language problem into a retrieval problem: pre-numbered, pre-tested prompts eliminate the randomness layer that comes from a model interpreting natural language
2. Three independent, orthogonal numbering systems for style/layout/color: 528 independently tested entries cover a theoretical combination space of nearly 2 million, drastically cutting maintenance cost
3. Multiple specialized working modes: from smart recommendation to precise selection, from poster design to batch illustration auto-fill, covering a range of real creative scenarios
4. An explicitly differentiated model capability matrix: deep-tuned models get direct output, general models fall back to reference images — not "one prompt set tries to serve every model"
5. A disciplined project scope: focuses purely on prompt asset management, not image generation itself, riding on existing AI agent infrastructure for distribution

Who This Is For

• Social media / official account content creators: need to quickly produce visually consistent cover images without improvising a style description every time
• Knowledge bloggers / infographic creators: need a stable, reproducible illustration style to maintain visual consistency across content
• Users already using Claude Code/Codex and similar AI agents: want to seamlessly fold AI image generation capability into an existing workflow
• AI art enthusiasts who've felt the pain of "prompt instability": want a validated, reusable style asset library

One-Line Verdict

HandRaw-Style doesn't try to make AI understand your description better — it simply removes the "description" step entirely, which might be the most direct way to solve the "style drift" problem after all.

Check out PrimeSkills — a curated marketplace of AI agents and skills that have been validated in real-world, enterprise-grade workflows. No fluff, just what actually works.

Find more useful knowledge and interesting products on my Homepage