Parrots are underrated pets and tools
I was glad to see this classic Theory of Constraints story called “Blue Light” by Kevin Fox on HN today. I had forgotten its lessons in the humdrum of life.
A consultant visits a welding plant running at “93% efficiency” with plans to expand the building. He has one simple heuristic: if the welding torch isn’t on, the welders aren’t welding. So he watches the floor.
I then watched the first welder begin to peel the protective plastic coating off the bumper in the places he had to weld. It took a good bit of time picking with his fingernails to get it done. Then he grabbed the parts and clamped them onto the bumper, put on his gear and welded for no more than 30 seconds before he was done. I looked at my watch, we had been there almost 5 minutes and he had welded for 30 seconds of it.
Everyone is busy all the time. Almost nobody is welding. The plant manager turns to the consultant and says:
“You see, they’re busy all of the time!” And he was right, the guys were working all of the time and working steadily and hard at that.
What amazed me is how the two of us could be looking at exactly the same things and see it entirely different. He had in his head an assumption of what good looked like that was based on the people being busy, whereas I looked at it from the perspective of the operation and the work it did, the blue light.
The fix was simple. They moved a summer intern into the department to handle everything that wasn’t welding. His one instruction: make there be more blue light. Within three weeks the backlog was gone.
Why can’t I just win all the time with heuristics
Section titled “Why can’t I just win all the time with heuristics”The blue light heuristic is trivially simple to describe. The bottleneck resource should be doing its core activity as much as possible. Measure the time the constraint spends on value-adding work, not support work. That’s the whole insight.
But knowing this heuristic exists doesn’t mean you’ll apply it when you’re staring at your own bottleneck. You’re too deep in the work. You’re looking at utilization numbers and capacity plans and thinking about headcount. The insight is in a blog post you read six months ago, not running in your head while you make decisions.
There’s a huge amount of known wisdom like this. Simple frames, outside perspectives, diagnostic questions. Most people encounter them once and never systematically apply them. You’d need a very good memory and spare cognitive bandwidth to keep all of it loaded while doing your actual job.
I think this is part of why smart people have better outcomes. They have bandwidth left over. They’re working on a task but they’re also running background threads, checking implications, pattern-matching against things they’ve read, noticing when something doesn’t fit. If you’re fully absorbed in the work, you get tunnel vision. You solve the problem in front of you and miss the frame-level mistake. “I didn’t think of that” is the common experience, but it happens less when you have more cognitive room.
Parroting is underrated
Section titled “Parroting is underrated”People like to dismiss LLMs as “just parrots.” I think parroting is underrated.
If you could take the blue light heuristic, encode it as a prompt, and run it against descriptions of your current bottlenecks, that would be useful. Not because the LLM understands manufacturing. Because it can mechanically apply a known-good frame to your situation and flag things you’re not thinking about. The value isn’t in the model’s intelligence. It’s in the heuristic, and the fact that the model applies it when you’d forget to.
You only have to think of the heuristic once. Then it goes into your rolodex and runs automatically. I don’t think people are using LLMs for this enough. The conversation is all about whether AI can write your code or replace your job. The more interesting use is: take known wisdom, encode it, and run it as a background process on your work.
I’m talking my own book here, but this is precisely why I’m working on Promptless. The specific application is documentation. Promptless watches your PRs, Slack threads, and tickets, and asks “does this change invalidate any existing docs?” That’s a blue light heuristic. Any tech writer could tell you to do it. But nobody runs that check in their head while they’re shipping code, because they’re absorbed in shipping code. So docs drift. Promptless just runs the heuristic continuously in the background.
Prosthetics, not replacements
Section titled “Prosthetics, not replacements”We all know people with more intellectual firepower than us. The coworker who sees implications you miss, the friend who wins the Putnam. Part of what makes them effective is they can run more parallel threads on the same task. They have room to check their work against a broader set of frames while doing the actual work.
LLMs are a prosthetic for that bandwidth. Not a replacement for thinking. A way to keep more threads running than your brain can hold. Your output stops being just the thoughts in your head. You get parallel checks for free.
Collectively as software enthusiasts I maintain we’re still barely scratching the surface. Promptless has been at it just over a year and it’s still a narrow application. But as models get cheaper and we get better at encoding heuristics in shareable formats, the ceiling is high.
What limits us as individuals and as organizations are the assumptions we hold, and our failure to recognize them as just that “assumptions” and not facts.