Use the cheap model for the boring 80 percent

1. Sort your day into two piles

Pile one: rewrite this, summarise that, what does this word mean, draft a reply, make a list, rename these files. Fast, low stakes, easy to check.

Pile two: plan this project, find the bug, work through this contract, compare these five options with real trade-offs. Slow, high stakes, hard to check.

Pile one is most of the day. Pile two is where being wrong costs something.

2. Put pile one on the cheap model

In the app, pick the small model and leave it there. Switch up only when you hit a pile-two job. You will be surprised how rarely that is.

I run my always-on agent on a cheap model for exactly this reason. Most of what it does is wake up, check something and decide nothing needs doing. That does not need a clever model; it needs a cheap one. When a job genuinely needs thinking, it hands off to a better model and then goes back to being cheap.

3. Let the model triage for you

Not sure which pile a job is in? Ask the cheap one first.

Before you answer: is this something you can do well, or should I take it to a more
capable model? Reply in one line, then do your best with it anyway.

If the cheap one says “take this elsewhere” and it is right, that is the system working.