Last night, I watched Liu Run's 900-Day Daily Lessons livestream on and off for two or three hours. One example stayed with me into today.

He said he greatly admired carpenters and electricians, the people we usually call repair workers.

When a repair worker enters a home, the situation is almost never exactly the same as in the last one. Different things are broken, the pipes and wiring differ, and so do the materials available. They inspect the site, take something from their toolbox, and use what the homeowner already has to take care of the problem.

“These people don't carry standard answers in their heads, but they have the ability to find an answer.”

Hearing that, I naturally thought about how I've been using AI over the last two months.

AI really has helped me make quite a few things I couldn't make before. Mini programs, images, short videos: with many things, once I have even a partial image in my head, I can try turning it into a prompt and have the model generate it. Especially since I recently began using Seedance 2.5 directly, the quality of the generated imagery often surprises me.

Surprise aside, there are plenty of problems.

When making Sanmu videos, it occasionally generates an extra hand for him. Speech, subtitles, and sound effects can also fall out of alignment: the image is at this point while the sound has already moved to the next sentence. Sometimes generating again fixes it; sometimes revising the prompt works. Other cases are stubborn. I repeatedly stress that something must not appear, and it generates it anyway.

When that happens, I no longer stay with the original scene and fight it endlessly. I keep what I wanted to express, change the setting, and try a different image. If I still can't find the cause, I ask people in the group who have gone deeper into it, and also give the result to AI to analyze where the problem lies.

What the video ultimately needs to accomplish is this particular expression. I don't have to cling to the first picture in my head.

The process resembles a repair worker visiting different homes. A method that works here may immediately fail in another setting. All you can do is look clearly at the problem in front of you, try an approach, examine the result, and decide what comes next.

I'm still far from proficient. I haven't gone deeply into shot planning, storyboards, character design, or voice consistency. Right now, I'm more at the stage of trying things lightly and feeling out the model's boundaries. More complex projects will certainly bring more problems. What I can accumulate is probably an understanding of why this attempt went wrong and how I later found a way around it.

“900 answers are waiting for a good question.”

That line also connects to the repair-worker example. The worker enters with accumulated experience; the homeowner's situation supplies today's question. The models, prompts, and production processes I've been learning are also adding to my store of answers. But answers have limited value if they stay in my hands. I need to meet real people and real needs, and hear a problem worth solving.

I haven't done this yet. I haven't bid for work or gained a new client as a result of this reflection. The idea has only just reached this point. There's no need to announce results ahead of time.

Next, I want to look at specific needs already appearing in the market. I could begin with a design request or a video commission. Or I could first sort out an unused camera or art supplies and carefully post one item on Xianyu. A real situation will tell me whether the pictures are clear enough, whether the copy speaks to what buyers care about, and what is still missing from my tools.

I'll still try new models, and I won't reject new tools. The desktop tools I currently invest most in are Kimi and Codex. With others, I'll start with the free versions. Once a tool can solve a specific problem and help me complete a real delivery, I'm willing to pay for it and invest more.

The old sequence

Look at the tool first, then decide whether to subscribe.

The sequence now

First see what's broken, then decide which tool is needed.

When a repair worker enters a home, they first inspect the situation. My next step is the same. Find a problem within reach of my abilities and try to solve it. The result may not be ideal, but at least the situation will leave me with new information.

“The toolbox will keep growing. I hope that, whenever I add a tool, I can say clearly what it has helped me fix.”