Ten Things I Learned About Coding with AI

1) AI isn't just a good thing, it's the only option.

Coding with AI can be much more productive than writing and debugging manually. It's a productivity multiplier and people who don't use it are going to have a competitive disadvantage.

2) Working with AI is like working with smart junior devs.

This is a skill to develop and refine. More-experienced development leaders have to be able to recognize and accept good solutions that aren't exactly what they originally imagined, but also understand when designs aren't quite right and know how to give instructions.  Treat it like pair programming with someone who is smart but inexperienced.

3) Some codebases work better with AI than others.

Even AI-generated code can grow to a point where AI doesn't work well with it. If it's maintainable by humans it should work well with AI and visa-versa. Osterhout's "A Philosophy of Software Design" is still relevant. Large, old codebases are going to need major refactors.

4) Use AI for code reviews all the time.

It will find things that you missed. It's not perfect - some of the things it finds will be unimportant or wrong but you have to figure out which suggestions aren't.

5) Code-review AI's work too.

I give Claude Code a task with guidelines and then let it run free. Then I test in a local environment with Claude reading the error log. It will find an fix many of its own errors. When it finally seems to work, I commit the code and only in the pull request go over every line that it produced.

6) Redesign your workflow to accommodate AI.

If you divide work up into small tasks with Jira, think less granular. Micromanaging will slow you down a lot. Pull requests will be much larger with AI. Velocity is through the roof, but not by conventional measurements

7) Product teams will become more closely integrated with development teams.

The instructions we give to AI are part product and part coding instructions and if you leave out one part of that it doesn't work as well.  Prompts are part product spec, part code spec. Include both for better results.

8) Just like junior devs, AI forgets things.

Put specific style and coding instructions in your CLAUDE.md or AGENTS.md (e..g, "Don't write code with N+1 queries" and "Use AJAX calls on the page to handle POSTs, PUTs, and DELETEs instead of using them to submit a page") and then follow up with your AI when it occasionally does them anyway.

9) Use AI to manage your TODO list.

If you use Claude Code, tell it to store the TODO list in a file so it won't be lost if Claude Code crashes (rare, but it happens) and it normally stores those lists in working memory.

10) Things are changing very fast.

Some of the problems with AI that I've described here will be fixed in future versions. I'm currently using Claude Code extensively, but experimenting with new products as people recommend them. Replit, Loveable, Codex, are all serious contenders and who knows what's next. But like humans, they all have personalities, strengths and weaknesses. Get to know your AI just like you would learn the personalities of your junior devs.