The AI does what you say, not what you mean

Julia is learning to build software this week. Her tool is Jules, Google’s coding agent: free while in beta, connects to your GitHub, reads the repository, works in a branch, and you merge when it’s done. It requires almost zero programming ▸ 12:11.

Which is exactly why the advice I gave her had nothing to do with code:

“El bicho hace lo que uno le dice. Pero no sabe lo que uno tiene en la cabeza. Y hay veces uno cree que lo que tiene en la cabeza es obvio, pero no es tanto.” The thing does what you tell it. But it doesn’t know what’s in your head. And sometimes you think what’s in your head is obvious. It isn’t. ▸ 16:08

The bottleneck moved

Before you start, have the idea fully formed: every feature, what happens on this click and on that one, described completely ▸ 15:53. The agent executes; it doesn’t fill the gaps in your intent. When the output is wrong, the spec was wrong: it just used to hide behind the months of coding that stood between spec and result. Now the distance is minutes, and vague thinking gets exposed at machine speed.

sé descriptivo o sufre →

This is the captain thesis from seed 1-1 growing its first ring: AI as the hands, the human as the head. The corollary we found teaching it: having a head is now the job. Clarity of intent became the scarce input.

The honest limitation

Jules has a real gap: it never shows you the program running. It debugs in its own virtual machine, but the problems you’d see in one glance at a screen only surface when you clone the repo locally and look ▸ 19:50. The hands are fast; the eyes are still yours.

Julia’s first project is already chosen, and it came from a restaurant queue. That story is planted as its own seed…