Yao Open Prompts Deep Dive: Turning Prompt Engineering into an Open-Source Craft
Project: Yao Open Prompts Author: yaojingang Website: https://yaojingang.github.io/yao-open-prompts/ GitHub: https://github.com/yaojingang/yao-open-prompts Scale: 116 Chinese prompts + 116 English mirrors Framework: RTF (Role-Task-Format)
The Core Problem
Prompts are the "translator" between humans and AI. With the same model, output quality can differ by an order of magnitude between good and bad prompts. Yet most prompt collections suffer from three issues:
- Fragmentation: prompts scattered online, often low-quality secondhand copies
- No structure: lists without classification, framework, or iteration history
- Hard to reuse: no guidance on adapting a prompt to your own scenario
- AI private-domain sales prompt V1.3
- Intelligent article polishing system V3.0
- PPT generator V1.0 → V3.0 (from basic generation to image-asset parsing)
- Webpage reverse engineering: extract visual, interaction, and technical patterns
- Image decompilation: analyze aspect ratio, fonts, colors, layout from a reference image (~90%+ fidelity)
- Article decompilation: deconstruct structure, style, and formatting into writing templates
- Video decompilation: break down short-video screenshots into prompts for text-to-video models like Veo 3
- Total prompts: 116 Chinese + 116 English = 232
- Highest version iteration: V3.0 (PPT generator, article polishing)
- Decompilation dimensions: 4 (web, image, article, video)
- GEO templates: 25 (full AI search optimization workflow)
- Content operations additions: 36 (short video, persona, platform ops, livestreaming, private domain)
- Framework version: meta-prompt system V0.6 (RTF)
- For everyday users: move from copy-pasting to understanding Role/Task/Format and adapting prompts to your needs.
- For prompt engineers: versioning, classification, and a shared framework turn prompts from an individual skill into a team capability.
- For AI application developers: for many AI apps, the moat lies in prompt quality, not the model. The repo offers commercially validated production templates—customer service (V1.0), Douyin viral content planning (V2.1), GEO content factory (V1.0).
- Decompile your workflow: break manual tasks (weekly reports, competitive analysis) into steps and extract "design genes"
- Build a team prompt library: organize internal prompts with RTF and version numbers following the repo's taxonomy
- Bilingual comparison: run the same request with Chinese and English prompts to see which suits your model better
- Yao Open Prompts website: https://yaojingang.github.io/yao-open-prompts/
- GitHub repository: https://github.com/yaojingang/yao-open-prompts
Yao Open Prompts addresses these not with "more prompts," but by making prompt engineering a reusable open-source craft.
Methodology
1. The RTF Framework: An Industrial Design Standard for Prompts
| Module | Purpose | Example | |--------|---------|---------| | Role | Define the AI's identity, expertise, tone | "You are a top short-video copywriter with 10 years of experience" | | Task | Clarify what to do and its boundaries | "Plan a topic system for the account, including virality assessment" | | Format | Specify output structure, style, constraints | "Output Markdown: topic list, rationale, expected metrics" |
Version V0.6 of the meta-prompt system adds two advanced layers: quality self-assessment criteria and iterative refinement based on feedback. The framework turns prompt writing from an intuitive art into a teachable, iterable engineering process.
2. Versioning: Prompts with a CHANGELOG
Every prompt carries a version number:
This makes prompts traceable, comparable across versions, and forkable.
3. The "Decompilation" Series: Extracting Design Genes
Instead of writing prompts from scratch, the project reverse-engineers excellent finished products into reusable RTF prompts:
This is pattern extraction, not plagiarism—the same logic as programmers reading open-source code to learn design patterns.
4. Nine Workflow-Based Categories
| Category | Count | Examples | |----------|-------|----------| | AI Work | ~20 | company research, contract drafting, customer service, PPT generation | | AI Content | ~36 | short-video copywriting, title crafting, article polishing | | AI Learning | ~10 | Feynman technique, keyword learning, mnemonics | | AI Methods | ~8 | meta-prompts, webpage/image/article/video decompilation | | AI Marketing | ~25 | GEO content factory, AI-friendly content, Schema.org optimization | | AI Education | ~6 | interactive learning pages, HTML5 mini-games | | AI Health | ~2 | personalized health reports | | AI Creativity | ~10 | Nano Banana image prompts, Sora2 video ideas | | AI Life | ~3 | parent-child songs, songwriting |
Classification follows human workflows—each prompt maps to a concrete task with a clear deliverable.
Key Numbers
Impact Assessment
Reviewer's Commentary (Feynman-style)
> Most prompt collections are hoarding—piling up whatever can be found online. Yao Open Prompts instead builds a standard for a craft. The RTF framework forces you to answer three questions: who are you (Role), what do you want (Task), and what output do I owe you (Format). 90% of people skip straight to "write me copy" without defining who the AI is or how it should respond. > > Versioning is the design I most want to praise. Most prompt collections are like code repositories without version control—everyone edits on master. The V1.0→V3.0 evolution paths show how a prompt iterates in real usage. > > The decompilation series has genuine hacker spirit: not "how do I write a prompt," but "how do I extract design genes from an excellent product." It turns prompt engineering from creation from nothing into pattern-based re-creation—easier, more reproducible, less likely to go astray. > > The GEO series shows real insight into AI trends. GEO (Generative Engine Optimization) is SEO's successor in the AI search era—when users ask ChatGPT/Perplexity instead of searching Google, "being cited by AI" is the new "ranking #1." Twenty-five GEO templates covering the full chain from opportunity assessment to compliance risk is early positioning, not trend-chasing. > > One weakness: 232 prompts can overwhelm beginners—the project needs a "starter path" (e.g., "if you're new to AI today, learn these 5 prompts first"). That's also a direction open-source contributors can take.