The new code is the prompt

Announced on camera as “the mega tip of how to build an agent, it’s incredibly easy” ▸ 24:48, and for once the hype label fits.

Cline, the VS Code agent this whole operation runs on, is open source. Open source means the interesting part is sitting in a public folder: the system prompt, the instructions that turn a language model into an agent. Going digging in the repository turned it up, and it is a beast, about a thousand lines of dense, carefully engineered instruction ▸ 25:31.

The observation that earns this entry its title: before, the valuable thing was always the code. Now the new code ends up being the prompt you give the agents ▸ 25:56, a series of instructions is what makes these things capable of doing everything you ask. The engineering moved up a layer, and in Cline’s case that layer is published for anyone willing to read it.

mil líneas, gratis, en una carpeta pública →

Then the practical move: take the thousand lines, hand them to Gemini, and say “help me adapt this for Structo”, specifying the operations Zenota’s agent needed, create, edit, move, color ▸ 26:21. Out came a robust agent prompt without writing it from zero, which is the quiet reason entry 21-2’s app shipped with a working assistant inside it so fast ▸ 26:36.

The general pattern is older than AI: when a craft is new and hard, find where the masters published their working notes. What’s new is what counts as the working notes. Nobody thinks to read a prompt the way they’d read source code yet. That gap is the advantage, for exactly as long as it stays a gap…