A regular person, learning AI from scratch.

I’m not a programmer by training, and not an AI expert. But I’m more and more convinced that AI is not just “one more tool.” It will change how ordinary people work, learn, organize experience, and build their own digital assets.

My daily AI usage is already a mix — ChatGPT for discussion and explanation, Claude for long writing and architecture calls, Codex for local execution and verification, Kimi for long Chinese context, local models for privacy and experiments. Each tool has its pros and cons in isolation; but once they all show up inside one real task, the problem isn’t “which is strongest,” it’s that the site gets messy.

A judgment lives inside a chat, a verification result sits in a terminal, a page state lives in a browser, a design decision is written to a local file. On the day, I obviously know how they connect. A few days later, all that’s left is a pile of material. The next session doesn’t necessarily know which one is a draft, which is mid-state, and which was an actual accepted result. What I’m missing isn’t more AI — it’s a place that can collect the experience.

I used to associate “lab” with very professional research institutions or flashy tech demos. My current understanding of a personal AI lab is more plain: it’s a long-term workbench. The workbench can hold tools and half-finished things, but the most important property is that it lets me look back — what I tried today, where I got the call wrong, why I switched methods later, which pieces are still worth keeping.

So YunLab.ai roughly records four kinds of things:

1. How AI tools enter real tasks — not benchmark numbers, but where each tool actually helped or where it was unreliable.
2. How agent workflows go from chatting to actually taking tasks — roles, boundaries, handoff, permissions, acceptance.
3. How a personal knowledge system goes from storing material to compressing judgment — chats, reports, screenshots, files only become assets after they’re organized.
4. How a long-term project like OpenClaw gets built one step at a time — the goal isn’t to show off results, it’s to leave judgments, mistakes, audits, and boundaries on the record.

If that process stays inside chat logs, it scatters fast. So I write the redacted part of my practice here: which ideas worked, which judgments were later overturned, which systems looked usable but weren’t yet trustworthy.

What stays here: methods, judgments, and stage retrospectives. What does not: private material, unredacted content, internal accounts, secrets, customer info, or project details that aren’t fit for the public.

So what you see is not the full backstage — it’s the slice of the experience I think is worth publishing.