Skip to main content
Phúc.
Back to Blog
AI2 min read

How I use AI to build software

How I bring AI into my day-to-day software work: understanding codebases, planning, writing code and reviewing — plus the principles I follow to keep AI genuinely useful.

AI does not replace a software engineer's thinking. But used well, it significantly changes how much time and how much energy I spend on each part of the job.

This post is a set of notes on how I use AI in my daily work — not a definitive guide, just what is working for me right now.

The tools I use

I don't chase every new tool. What I need is an assistant that works directly on the codebase: reading files, running commands, making edits and checking the results. So my main tool today is Claude Code running in the terminal, next to my usual editor.

Where does AI fit in?

1. Understanding a codebase

When I approach an unfamiliar module, I let AI read it and summarise the main flow, the dependencies and the risky spots. It does not replace reading the code myself, but it tells me where to start reading.

2. Planning before coding

For larger changes, I ask AI to propose a plan: which files to touch, in what order, the risks and how to test. I review the plan the way I would review a colleague's design — trimming and adjusting before any code is written.

3. Writing code in small steps

I break the work into verifiable steps. Each step ends with a type-check, lint or test run:

npm run lint && npx tsc --noEmit && npm test

Small steps let me catch it early when AI heads in the wrong direction, instead of untangling one big change.

4. Review and bug hunting

AI is a decent "second reviewer": it spots missed edge cases, missing error handling or code that could be simpler. The final call is still mine.

Principles I follow

  • Context is everything. Describe the goal, the constraints and the project's conventions clearly. Good results come from good requests.
  • Never merge what you don't understand. If I can't explain a piece of code, I'm not ready to own it.
  • Verify with tools, not feelings. Tests, type-checks and actually running the code beat an answer that merely sounds right.
  • Protect data. No secrets, customer data or sensitive internal information in prompts.

Conclusion

AI lets me spend less time on repetitive work and more time on what matters: understanding the problem, designing the solution and ensuring quality. I'll keep logging concrete experiments in the Lab.