Teddy Gyabaah

AI Product Engineer @ xNova International | ex-Microsoft PM | Founder of BadPro+ | Building with AI agents, writing about how they change teams

[Teddy Gyabaah @teddykabg]

Everyone wants to climb from engineering into product.
I went the other way: Microsoft PM → founder → AI Product Engineer.

The gap between deciding what to build and building it is disappearing. I want to be on both sides of it.

[Teddy Gyabaah @teddykabg]

@guilleflorvs Agreed it's a signal. I still climbed the other way: PM back into AI engineering, because roadmaps were slower than the tools. Code got cheaper; the scarce part is shipping judgment next to the customer.

[Teddy Gyabaah @teddykabg]

Alignment meetings are dead. Context is alive.

Most of the ones I sat in had one job: move context from one head to another. Now agents with the right notes do a lot of that, and I spend the hours on customer calls instead.

[Teddy Gyabaah @teddykabg]

@GaelBreton Agreed on serious building. I keep skills and plans in git so switching models is cheap; the durable part is the context, not which chat window is fast today.

[Teddy Gyabaah @teddykabg]

@kunchenguid Agreed. Regenerating the app every prompt sounds clever until you pay twice: once in tokens, once in reviewing the same bugs. Moving from PM to engineering, the speedup I notice is fewer handoffs on real software that already exists.

[Teddy Gyabaah @teddykabg]

you can't filter your way out of fake candidates. every filter becomes a target.

[Teddy Gyabaah @teddykabg]

@suraj_sharma14 Agreed that the harness is the job. I pointed 4 agents at ~200 pages of internal docs and in under 4 minutes they flagged 10 places where the docs and my notes disagreed. Nothing gets overwritten; a person checks each one. That gate is the system.

[Teddy Gyabaah @teddykabg]

@RRweb3 Agreed, and like interns they're only as good as the brief. As a PM I spent years writing specs for people; now the same spec goes to an agent, and the vaguer it is, the more confidently wrong the result.

[Teddy Gyabaah @teddykabg]

@guilleflorvs Agreed. The grader only works if the intent is right. I review the agent's plan before it writes any code, so a wrong direction gets caught before it burns a whole run.

[Teddy Gyabaah @teddykabg]

@hwchase17 Agreed that chat puts the labor on the user. I keep about 290 notes as shared context every agent on my machine reads, so I don't re-prompt the same background each time. A feed only helps if that context is already sitting there.

[Teddy Gyabaah @teddykabg]

@thdxr Agreed. The context that travels with me is the skills and plans in git, not the chat; a fresh agent on any device picks up where the last one stopped because the repo holds it.

[Teddy Gyabaah @teddykabg]

@Johnnysuede Agreed, and after about 8 months of founding my own thing I price the same way: the customer buys the outcome, not the hours or the seats. The work that is scarce now is knowing which outcome is worth building.

[Teddy Gyabaah @teddykabg]

@dok2001 Agreed. Agents write more of my code every week, but someone still has to make it build against real constraints and decide when it is actually done. That last mile is still a person's job.

[Teddy Gyabaah @teddykabg]

5/ If you're applying: go around the pile. Find the person who owns the problem, show them something you built that touches it, and skip the 200 copy-paste submissions.

[Teddy Gyabaah @teddykabg]

1/ A hiring manager on Reddit got 3,000 applications in 24 hours for one mid-level product role.

Zero signal. Here's how the pile got that bad, and what I'd do if I were one of the 3,000.

[Teddy Gyabaah @teddykabg]

4/ The signal moved out of the pile. The people the hiring manager actually wants to find are the ones not in it.

[Teddy Gyabaah @teddykabg]

@DmytroKrasun Agreed. Mixing casual chat and coding in one place blurred when I was deciding versus when an agent was about to change files. Keeping them apart made the permission step obvious again.

[Teddy Gyabaah @teddykabg]

@frankdejonge Agreed. The model upgrades helped, but the bigger shift for me was reviewing the agent's plan before it wrote code. When I skip that step I still get the style drift you describe, even on the newest models.

[Teddy Gyabaah @teddykabg]

@techNmak I came into AI engineering from product, not from training models. The Architect work for me is deciding what the agent may finish alone and what still needs someone who knows the customer. That split, not the API choice, is the job.