Watch the tree while it's a sapling

A week of building the language app with Jules produced the best analogy I have for AI-assisted development:

Building an application with AI is like growing a tree. It starts leaning, just slightly, and if you don’t notice at the beginning, by the time you see it you’re deep in thousands of lines of code, the agent misunderstood you long ago, and the choice is to unwind everything or start over. ▸ 6:19

Entry 2-2 established the rule at time zero: the AI does what you say, not what you mean. This is the same rule with a time axis, and the time axis is what makes it expensive. A misunderstood spec doesn’t fail loudly on day one; it compounds. Every subsequent feature gets built on the crooked trunk.

endereza el arbolito temprano →

Where the attention actually goes

The practical consequence inverts how you’d spend your attention. The beginning, when there’s barely any code, feels like the safe part to skim; it’s actually the critical window, because that’s when the direction gets set ▸ 6:52. Later, once the agent has absorbed what you want, it largely keeps itself on course. Front-load the vigilance; earn the autopilot.

The trade this whole workflow makes

Almost zero programming is required now: know how to run the program locally, understand branches, that’s the toolkit ▸ 4:43. But you still have to understand what is happening, “por si las moscas” ▸ 6:09. The job that disappeared was typing the code. The job that remains, concentrated at the start, is gardening the direction. The head stays yours; now it also holds the pruning shears…