AI Agents and the DIY Side of Software

Generative AI and AI agents have shortened the distance between having an idea and starting to build software.

Not long ago, noticing a small inconvenience in daily life was one thing; turning it into software was another. You first had to learn a fair amount about design and programming.

Now I can talk an idea through with an AI and, even in a short session, turn it into something that actually runs.

It is not like building furniture from scratch. It is closer to putting up a shelf, or adjusting something so it works a little better.

To me, it feels like the DIY side of software — weekend carpentry, more than engineering.

From talking things over to working together

As I write this, I mainly use ChatGPT and Claude Code.

ChatGPT is where I talk through ideas, rough specifications, and screen layouts. It also helps me put vague thoughts into words, and sort the features I need from the ones I don't.

Claude Code reads the actual project, adds and edits code, and checks that things run.

Until fairly recently, generative AI was something that answered questions with sample code.

Today's AI agents read multiple files, rewrite code, and run commands as they work. Instead of copying answers out of a chat window and applying them one by one, I share a workspace with the AI and we work on the same code together.

Larger software is still too much to hold

That said, generative AI does not mean anything can be built with little effort.

As features grow, so does everything that has to be thought about: design, screens, data handling, security, and what happens after release.

Large software is more than one person can hold. And for now, the same is true of generative AI and AI agents.

When I hand over too wide a scope at once, parts that should not change get changed — or each fix works on its own, while the whole quietly loses its shape.

What to build.

How much to delegate.

Whether the result is safe to release.

Those calls are still made by a person.

Making daily life a little easier

So for now, I build small things I can use right away.

Something that makes me pause before sending an email or a message.

Something that hides what I typed.

Something that shows what information is still left inside an attached file.

Life goes on fine without any of them.

Still, having them around lets me feel a little safer, or skip one tedious check. That is the kind of tool they are.

I don't keep them small because they are quick to build. I keep them at a size where I can understand the purpose myself, verify how they behave, and explain them to the people who use them.

Generative AI and AI agents have made this kind of software DIY much easier to try.

A few years from now, the way software is made — and how the work is divided between people and AI — may look very different from today.

Legacy Tools is, in part, a record of the small tools I started building along the way.

Written: July 2026. Tools in use: ChatGPT (GPT-5.6) and Claude Code.