I’ve noticed something recently, tasks where I explicitly tell the agents **how** to perform them end up way better than the ones where I just tell them **what** to do.

![](/media/01M43VZQANAJT8V5ZR22TZNENH/tell-the-ai-how-not-only-what-1.png)

The idea clicked last week when I asked my agents to build a Single Page Application to perform a hypothesis test, i.e. tell me if some experiment we were doing had a noticeable difference or if it was just noise.

I asked for a probability instead of a p-value, and that very naturally produced a Bayesian Statistics solution. But I was very suspicious of the results. When I investigated a bit more, the solution was very sound, but overly complex and not quite right.

I tried again, with small nudges. It went *worse*.

I let the idea simmer a bit in my mind. I find it very useful to “sleep on a problem” (better than sleep, a long walk works wonders).

I started to remember every time that happened and compare those cases with the times when I had absolutely no problem. The difference seems to be whether I explicitly say **how**, not only **what**.

That also checks out when I tell the AI to implement some code. I’m usually very explicit about the architecture and component reuse, and that goes very well.

Which means it’s time for round 2 (more like 3 or 4 at this point 😮‍💨).

Warning: statistics gibberish ahead.

I specified that I wanted it to use *Bayesian Bootstrap* for a *paired test* and give me the *net improvement* with the *credible interval*. I also specified how I wanted to visualize the data and, since the Single Page Application would use TypeScript, told it to check the results against reference code in Python using scientific libraries.

That bunch of statistical nonsense above describes what I wanted **and** how to make it. I honestly don’t need to explain too much to the AI. If I use the right high-level terms, it knows what to do. The trick is that **I** also need to know the “how” in high-level terms and what that’s supposed to produce.

Guess what? It went exactly as intended.

Which makes me conclude the following:

- Telling the AI “how” removes a lot of sloppy, wrong guessing. The results are just better.
- A high-level “how” is good enough. I don’t even need skills for things the AI already knows how to do.
- The human in the loop here *still needs to know the how*. I don’t need to remember the formulas, but I need to know what they are, why and when to use them, and the caveats.
- The good part is that I can use AI to help me understand that better 😉

Which makes me think that’s the new role for jobs heavily affected by AI: instructing agents on what to do **and** how to do it.

We don’t need coders anymore, but we still need software architects.

Maybe?