Here are some lessons learned from AI Coding
1. Plan a lot
Claude Code feels so magical that you just want it to do everything fast. Back in July and August, my prompt to CC was simple: “Can you build a main menu for the app?” The result? It started inventing things I never asked for.
By September, I’d learned a bit. I wrote: “I would like you to write a plan for a main menu for the app.” Better, but still too vague.
Now, the prompt looks like this: “I’ve attached a short plan for a main menu for the app. Can you review it and suggest a detailed task list that I’ll review?”
That’s been the real lesson: the magic of AI isn’t in giving it freedom. It’s in learning how to speak its language.
2. Sometimes AI is so lost
My app, ExamGenie, turns photos from school textbooks into gamified learning content using AI prompts. I’ve been iterating endlessly to find the right prompt that can produce everything needed. Often, the quality of the exam just isn’t there.
CC suggested adding more and more guardrails, making the prompt balloon to over 200 lines. I had to say, hold on a second. I asked if such a massive prompt actually works better than a short one that lets the LLM act more like a teacher. It hadn’t considered that but instantly agreed I was right.
Sometimes, it just feels a bit dumb. It struggles with creative, out of the box thinking. Instead of stepping back and trying something new, it keeps zig zagging down the same path, fixing problems in the same way. It’s not good at breaking the pattern.
3. ChatGPT & CC = sometimes you need both
There was a big problem getting the Flutter app to run properly on iOS. I ended up starting a fresh project to isolate the issue, but got stuck again with CC on camera initialization (the app uses the camera to take images of school textbook pages). Endless loop of trying things out, nothing works.
So I went to ChatGPT. I brought along a document that CC had written about the problem to give ChatGPT full context. I asked for creative solutions, and it immediately suggested ideas that CC hadn’t even considered.
I think there’s real value in using two different LLMs for solving tough technical problems. They approach things differently, and that’s often where breakthroughs happen.
More lessons tomorrow.
