A prompt cookbook for when you want AI to look human

What it does

This is a collection of 19 validated prompt recipes for hand-drawn image generation. You tell an AI agent what to draw and which style number or alias to use—bean, ink-wash, softnose, etc.—and it returns a clean, copy-pasteable prompt. The project stays strictly out of the image generation pipeline; it only produces text prompts or, for one advanced style, a structured JSON call package with anchor images and a multi-stage workflow.

The interesting bit

The real work isn’t the prompts themselves, but the protocol layer that keeps agents from “freestyle-drifting.” A Python renderer script extracts recipes verbatim to prevent an agent from abbreviating or remixing the style descriptions, and one style (3.1) even ships with a SHA-256-verified reference anchor and a mandatory three-stage correction workflow. It is essentially a style contract between you and your agent.

Key highlights

• 19 locked-down styles spanning eight categories: realistic hand-drawn, line art, “intentionally bad,” retro, animation concept, paper craft, storybook, and 3D vinyl toys.
• Tool-agnostic core: two markdown files (PROTOCOL.md and STYLES.md) with thin adapter layers for Claude Code, Cursor, Codex CLI, Gemini CLI, and others.
• A Python renderer (scripts/render_prompt.py) extracts prompts exactly as written, avoiding agent drift during “stable production.”
• Style 3.1 (rawkid) includes a formal JSON output with reference image anchors and a forced three-stage workflow for consistent output.
• Proportions are soft-structured rather than hard-locked; the agent respects your aspect ratio if you provide one, and only injects layout hints (like “vertical multi-panel”) for comic styles.

Caveats

• Style 3.1’s “stable production” mode requires a tool chain that can parse JSON call packages, handle reference anchors, and execute a forced three-stage workflow; it is not a simple text prompt.
• The README cuts off mid-sentence in the contribution section, so extending the library with your own validated styles may require reading between the lines.
• Value is tightly coupled to agent-based workflows; without an AI agent that reads custom instructions, you are mostly looking at curated markdown.

Verdict

Worth installing if you live inside Claude Code or Cursor and generate a lot of stylized diagrams, storyboards, or concept art. If you already have a personal prompt library or don’t use agentic coding tools, this is just very well-organized markdown.