AI Talk - Code Documenting and Testing with AI
Automated testing is the job everyone agrees matters and almost nobody gets done. It's slow, it's boring, and the people who are good at it are expensive. So on a codebase that's been growing for years, it keeps sliding to "next quarter".
Pedro Dobrescu tried to start it at least three times in seven or eight years on the same project, a ShortPixel codebase. One attempt, with a part-time specialist, took six months to produce a setup and not a single test, and then got thrown away. With Claude Code agents, 80-90% of it was done in two weeks: 2,000+ unit tests, 400+ integration tests, 80+ compatibility tests, ~250 browser tests, and smoke tests against the live APIs the project depends on; that last one was the agent's idea. Pretty much every product feature is now tested end to end. Pedro isn't a QA specialist. He's a generalist, and that's the point of this session: what someone without the expertise can now do, and what he learned the hard way.
What We'll Cover
The story: struggle, struggle, boom
Three failed attempts at automated testing, and why they failed
What changed when agents came in
Document first, then test
Getting the agent to document the whole codebase first (doc blocks on every class and function) and the bugs that turned up just from that sweep
The testing ladder, piece by piece: unit tests, integration, compatibility, platform-specific, then browser tests that behave like a real user
Live: how to start
A project with no documentation and no tests, opened in Claude Code's planning mode
The prompt, the sub-agents, and the plan it comes back with, the starting point you can copy
Open Discussion
Your stack, your codebase, your testing gaps; every project is different, so bring yours
You'll Leave With
Why an automated test feedback loop is what lets an AI agent actually close the loop: without it, it can't reproduce a bug or check that its fix works
A starting sequence you can reuse: document, then test, one layer at a time
How to kick it off with planning mode and sub-agents, instead of one endless chat
An honest one: do this early, before the project gets years ahead of you
Who Should Come
Developers and tech leads on a codebase that's been around for years and has few or no tests
Product people and founders who own a product and keep hearing "we don't have tests for that"
QA and UAT people curious what agents can take off their plate
Generalists who code a bit and want to know where to start
Your Hosts
Pedro Dobrescu: product manager since 2005 and entrepreneur since 2011, with product as his main focus and some coding on top.
Ady Rugina: 20+ years of international product experience at companies like eBay, Orange, and Just Eat Takeaway. Creator of the AI Talk series and part of the organising team at Product Makers Romania.
Andreea-Ruxandra Dumitru: part of the AI Talk organising team at Product Makers Romania.