I spent the last three months using an AI-first editor. At first, it felt like a superpower. I could hit a keyboard shortcut, describe a feature, and watch 50 lines of React code appear. It felt fast. Then, the bugs started hitting.
I realized I had developed a dangerous habit: I was accepting code I did not fully understand. When a production bug appeared in a component written by the AI, I spent forty minutes trying to figure out the logic because I had not actually written it. I was the reviewer of my own code, but I was a lazy reviewer.
Here is how I changed my workflow to actually stay in control.
The Trap of the Big Prompt
The biggest mistake is asking for a whole feature in one go. If you prompt "Build me a paginated table with sorting and filtering," the AI will give you a massive block of code. It will look correct. It will probably work on the first run.
But that code is a black box. It often includes unnecessary dependencies, weird edge-case handling that doesn't fit your project, or outdated API patterns. When you accept a 100-line diff, you are essentially signing a contract to maintain 100 lines of logic you did not architect.
The Atomic Prompting Method
Instead of big blocks, I moved to atomic prompts. I break the task down into the smallest possible logical units.
First, I define the types or the interface. I prompt for the TypeScript definition only. I review it. I tweak it. Once the data shape is locked in, I move to the next step.
Next, I prompt for a single helper function. For example, "Write a function to sort this specific array of objects by date." I test that function in isolation.
Finally, I ask the AI to wire these pieces into the UI. By the time the component is finished, I already understand every helper function and every type definition because I guided the process step by step.
Using the AI as a Rubber Duck, Not a Ghostwriter
I started using the chat feature differently. Instead of saying "Fix this bug," I now say "Explain why this bug might be happening and give me three possible directions to investigate."
This forces my brain to stay engaged. If the AI just gives me the fix, my brain switches off. If the AI gives me the theory, I still have to do the work of implementing the solution. This is where the actual learning happens. If you stop solving problems, your skills will atrophy. You become a prompt engineer, which is a fancy way of saying you are a middleman between a requirement and a machine.
The Code Review Rule
I implemented a personal rule: I cannot commit any AI-generated code until I can explain every single line to a hypothetical junior developer.
If there is a regex pattern the AI generated that looks like magic, I do not accept it. I ask the AI to break down the regex. I test it against a few strings. Once I understand the magic, I commit it.
This slows down the initial development speed, but it drastically reduces the time spent in the debugging phase. Speed is a lie if it leads to technical debt that you cannot navigate.
Concrete Takeaways
If you are using an AI editor, try these three things this week:
- Limit your diffs. If an AI suggests more than 20 lines of change at once, break your prompt into smaller pieces.
- Prompt for the "Why" before the "How". Ask for the logic or the plan before you ask for the code.
- Manual Refactor. After the AI generates a working solution, manually rename variables or restructure a loop to fit your style. This forces you to read the code carefully.
AI is a great tool for removing boilerplate, but it is a terrible architect. Keep the architecture in your head, and use the AI to fill in the gaps.
Top comments (0)