Know your AI built exactly what you asked for.
I help senior developers ship AI-generated code to production with confidence.
I'm Girish Patel, a software engineer in the UK. I've been building commercial software for 22 years and practising test-driven development exclusively for the last 10.
I've turned that into a workflow for using AI coding agents on production software without handing over control of what gets built. It's for experienced developers who already use AI to write production code and want evidence that it built what they asked for.
The idea, on one screen
You and the agent agree the scope as a few plain-language scenarios, and you sign them off. The agent builds them test-first from the outside in. Then mutation testing runs on the lines that changed, and anything no scenario asked for is deleted.
Scenario Exporting an empty report
Given a report with no rows
When the user requests a CSV export
Then they get a file containing only the header row
survived src/export.ts:42 if (rows == null) return []
No signed-off scenario needs this. Deleted.Read the whole flow. It's free, and it will stay free.
Writing
- Use mutation testing to delete code, not to score tests
An AI agent will happily write more than you asked for. Mutation testing, pointed at the right question, is how I take it back out.
The course
The method is free. The course is for when you run it on your own codebase and it doesn't go to plan: practice repos with planted defects, a way to score your own results, and a guide to what went wrong. What's in it.
Want to talk?
If you use AI to write production code and want to talk about how you check what it builds, get in touch.
The call is free. I'm using these to learn what people get stuck on.