Automation
- Input
- Fixed Prompt
- Output
- Repeat
Same rules.
Different input.
AI Self-Evolution Experiment
AI Self-Evolution Experiment
An AI Designer That Learns to Improve Its Own Taste
一个会持续审视自己的设计,
调整设计规则,并逐渐形成更好审美的 AI 设计师。
The goal is not to generate more.
The goal is to learn how to generate better.
EvoDesign 是一个关于 AI 自进化的设计实验。普通 AI 自动化通常使用固定 Prompt,每天处理新的输入,然后持续产生新的输出。
EvoDesign 想探索的是:如果 AI 每次完成设计之后,还能重新审视自己的结果,判断哪里做得不好,并修改下一次执行任务的方法,那么长期运行以后,它是否能够逐渐形成自己的设计原则与审美?
变化的不只是 Output,
而是产生 Output 的方法。
同样会产生结果,但只有其中一种会改变产生结果的规则。
Same rules.
Different input.
The system changes
how it works.
一张关于未来运作方式的概念图。它描述方向,而不是当前已在运行的系统。
Concept model only — these artifacts are future design goals.
Traditional automation updates data. EvoDesign updates the rules that create the result.
A compact record of what this first phase is — and is not.
| Item | Description |
|---|---|
| Project | EvoDesign |
| Category | AI Self-Evolution Experiment |
| Current Phase | Phase 1 — Concept |
| Core Object | Design Prompt |
| Input | Previous Design + Reflection |
| Future Output | New Design Generation |
| Evolution Target | Prompt + Design Rules |
| Future Trigger | Daily Cron Job |
| Evolution Evidence | Prompt Diff / Design Diff / Logs |
| Website | Cloudflare Pages |
| Current Status | Concept Only |
Concept Preview
Not real execution data.
“Yesterday's design was structurally clean, but relied too heavily on cards.
The layout was easy to understand, but visually predictable.
For the next generation, I would reduce container usage and rely more on whitespace, typography and spatial hierarchy.”
Before
After
Prefer spatial grouping.These are not performance scores. They are a possible vocabulary for describing current design preferences.
真正的自进化必须可以被证明,而不是只说“AI 变聪明了”。
What changed in the prompt?
Why did the AI decide to change?
What changed in the design?
When did the change happen?
Evolution must be observable, comparable and traceable.
Verified Runs
Recorded repository evidence, not concept preview data.
006
v1.0
28 September 2026
CHANGE
Live evidence changed from a first-run snapshot to a generation history.
Identified that the evidence area did not scale with recurring generations.
Recorded the need for a durable generation-history surface.
Observed that the live status pattern could not absorb multiple generations.
Observed scattered status language instead of scalable history.
Added the first verified live-run evidence boundary.
Latest Reflection
A recurring experiment needs a dated evidence sequence, not a static first-run card.
Latest Change
The verified evidence surface now summarizes Generation 006 and shows prior repository-backed runs.
The experiment now has a visible trail.
It can be inspected over time.