aipat.chDarren Sharman
Case studies

From idea to phone

Making software development more productive

How I build with AI agents, what it has come to since March 2026, and an app port that took about an hour from asking to running on my phone.

projects since March 2026
76
commits
6,554
automated tests
20,015
from asking to an app port on my phone
About 1 hour

The situation

In March 2026 I started building with AI agents full time. This site is me exploring what that means for me, and hopefully for other software engineers. I think productivity will go up, but the harder problems are still going to be harder problems, and you still need someone in the loop who knows how to deal with the code coming out of the AI. Here that person is me.

What I did

I set the agents up a bit like a software team. Each project has its own Claude, working in its own folder from a spec that is the source of truth. I tell Control what I want, usually by voice. Control knows every project, and it starts the work in the right one. Anything with a UI or an open UX decision gets 3 to 5 working throwaway prototypes first. I pick one, and the spec describes the winner.

  1. Asking

  2. Planning

    • Spec the source of truth
    • Prototypes 3 to 5, then I pick
    • UX and UI bar a measurable floor
  3. Doing

  4. Proving

  5. Delivering

Always on

Tools in teal have their own project page.

Every Claude starts with the same rules, injected at session start. What it learns gets written down in docVault, and Lore keeps every conversation. Nothing counts as done until it's proved. Test Gate blocks a push if the number of tests drops, the e2e testers walk the app like a user, and Quality scans for the mistakes AI tends to make. Findings that keep recurring become rules.

The result

Since March 2026 that has come to 76 projects, 6,554 commits and just over 20,000 automated tests, with me as architect and product manager and the agents writing most of the code. Much of it is plumbing, and a lot of it only has one user.

One timed example. Transcriptor takes a video link and gives back what Claude made of it. On 30 October at 16:32 I asked for a Kotlin version of its Android app. The replacement was committed about forty minutes later, and at 17:38 I'd signed in on the phone and the library looked right. About an hour.

What a team can take from it

What has made the difference for me is the checks around the agents. The prototypes, the tests and the walk-through on a device happen every time, because an agent does the legwork and the gate won't accept the work without them. I think a team bringing AI in would get the most out of putting those checks in place first.

If you'd like to talk about this, get in touch.

More case studies