There is a right and wrong way to do Harness Engineering.
What makes it right or wrong mostly comes down to whether you’re following the lessons of Bitter Lesson Engineering or not.
Bitter Lesson Engineering comes from Richard Sutton’s “Bitter Lesson” essay, and it means ensuring that you’re not trying to outsmart your own AI. It means not trying to micromanage how your AI does things, but rather specifying what you want done.
Plainly stated:
- Bad Harness Engineering is a whole bunch of prescriptive instructions on exactly how to do things.
First copy this file, then load this, then do this, then do that. Etc.
- Good Harness Engineering is about providing tons of context about who you are, what you’re about, what you’re working on, what you’re trying to accomplish, and what good (and bad) look like to you.
I’m an engineer focused on personal productivity, I like simple designs with lots of whitespace and great typography, here are my previous projects, here are some tools you can use, etc.
Bad Harness Engineering is bad because the smarter AI gets the more antiquated your instructions will become. And at some point (maybe even now?) they’ll make your AI stupider instead of smarter.
Good Harness Engineering is good because no matter how smart an AI becomes it will still be better at getting you great results if it understands who you are and what you like.
Basically, both your prompts and your harness should be about who you are and what you’re trying to accomplish, and not specifically how to get there. That’s what the AI is there to figure out.
Give it the best possible picture of you, your ideal outcome, and the best tools you can, and give it room to work.
