The Matrix trick, load the skill not the manual

The naive way to give an assistant a lot of expertise is to write it all down and paste it into every session. One giant instruction file, loaded up front, every time. It is expensive, and it buries the one paragraph that matters today under the dozens that do not.

The image that fixed it for me is from The Matrix (credit to the Wachowskis): Neo learning kung fu in a few seconds, because the skill loads the instant it is needed and not before. The assistant should stay mostly empty. A task calls for a skill, that skill’s instructions load, get used, and fall away again.

That is how Sapix handles its skills. They sit dormant, dozens of them, and at the moment of need it recommends the one that fits, and only that one enters context. The cost tracks what you are doing now, not everything you might ever do.

Two layers fall out of this. The always-loaded file shrinks to a bootstrap: who am I, where do I look, the few rules too important to ever miss. Everything else lives in the vault and gets recalled on demand.

A smaller thing I only learned by doing it: for this, a skill plus a little command-line tool beats a skill plus an always-on server (the kind you wire in over MCP). The server pays rent every session whether you use it or not. A tool that runs only when invoked keeps the resting cost near zero, which is the entire point.

None of this is clever, honestly. It is mostly refusing to fight how a mind already works. You do not hold every skill you know in conscious attention at once; you recall the one the moment asks for. Context is a working set, not a warehouse.