Giving the model hands
A language model, for all its brilliance, can only produce text. A tool is how you let it reach through the glass and act — and once you see it, the ceiling on what you can build quietly lifts.
Longer thoughts that wouldn’t fit in a single line. Written slowly, edited slower.
A language model, for all its brilliance, can only produce text. A tool is how you let it reach through the glass and act — and once you see it, the ceiling on what you can build quietly lifts.
For more than fifteen years my moat was knowing how to speak to computers. Today I watched a room of engineering leaders admit, without flinching, that their teams have stopped hand-coding. I want to mark it.
Designing with AI is a solved problem. Getting a coding agent to honor every pixel of that design — without giving up halfway and calling it done — is the new craft, and I think it’s a two-year differentiator.
We invoke coding agents with prompts and skills and assume the instruction is the outcome. It isn’t — and the gap is where reliability lives.
My agent’s memory kept growing until it filled the very room it was meant to think in. The fix was to stop loading it.
If you live inside AI tools, you’ve been riding a world-class agent — you just never wrote the loop. Here’s the whole thing, and it’s smaller than the noise.