Prompt engineering, learned the honest way: by failing at it

What it does

This is a long-form, personal exploration of prompt engineering by phodal, written in early 2023 as the AIGC wave hit. It walks through three creative domains — text-to-image with Stable Diffusion, article writing with ChatGPT, and code generation with GitHub Copilot — showing how prompts evolve from vague descriptions into precise, structured inputs. The repo is mostly a Jupyter-notebook-backed essay, plus a companion tool (click-prompt) that packages the patterns it describes.

The interesting bit

The author’s framing is refreshingly blunt: AI is like Captain America — strong against strong opponents, weak against weak ones — and the prompt is the whole game. The image-generation section is essentially a documented failure log: verbose ChatGPT-embellished descriptions get stripped down to keyword-style prompts, then rebuilt with negative prompts, ControlNet for pose control, and DreamBooth for personal style. There’s also a speculative essay arguing “prompt programming” is the real no-code — code as a disposable byproduct of natural language.

Key highlights

• Covers the full Stable Diffusion prompt pipeline: negative prompts, ControlNet, DreamBooth, and inpainting for hand repair
• Includes a DDD + ChatGPT + Copilot worked example for code generation
• Companion notebooks for translation and code generation exercises
• Frames a personal AI strategy: strengthen architecture skills, build domain-specific small models, keep experimenting
• Honest about the author’s own limits — explicitly not a deep learning expert, writing to update his own mental model

Caveats

• Written in Chinese, and dated February 2023 — the tooling landscape it describes has moved considerably
• The author warns some image sections contain age-restricted content
• It’s an essay with notebooks, not a library — expect reading, not installing

Verdict

Worth a look if you want a practitioner’s field notes from the early prompt-engineering era, especially the image-generation workflow. Skip it if you’re after current, production-grade prompting techniques — this is a historical document, and a charmingly candid one.