If it's a legacy/unknown codebase, it's a good idea to research it by asking questions to Cursor/Claude Code/Codex
Then it's important to have minimal testing to validate that there are no errors in the code (lint, typecheck) after AI changes. Some agents will need rules, others will always try to do some testing.
It's vital to test every code change manually, so having local debugging tools setup is also very valuable (scripts, commands, ui/browser interactions). This might be skipped in the future as iterations are faster and safer/bug-free.
Asking for the feature itself really depends on each engineer's experience. Usually, describing the feature as much as possible with voice transcription tools does the trick. Using "Plan" modes + "make me questions" or "interview me" is another good trick for getting more context and planning.
Sometimes, just trying out a one-shot solution is also a good way of starting a feature since we can get misconceptions really quick.
Prototype first, design it later.
Top comments (0)