Tokenmaxxing
I have been saying this word to people for three weeks and getting blank stares and I need to write it down so I can just send a link.
Tokenmaxxing. It is the thing. And almost nobody is doing it intentionally even though the people who are doing it are kind of running laps around everyone else right now.
Here is what I think happens. You get access to a good model. The model is genuinely impressive. You ask it stuff, it does stuff, sometimes it is great and sometimes it is whatever. You conclude that it is inconsistent. You move on.
The model was not inconsistent. Your inputs were.
Tokenmaxxing is just "okay brace yourself" caring about what you put in as much as what comes out. That is it. That is the whole thing. I know. I know it sounds obvious when I say it like that. And yet!!! Look around!!! The median AI workflow is "type thing, hope, paste into Slack, ask if it seems right." We are out here handing a chef random unlabeled ingredients and being surprised the dish is weird.
The contractor analogy
Imagine you hire someone brilliant. Like genuinely talented. They show up day one and they have never seen your codebase, do not know your constraints, have no idea about the thing Dave did to the auth layer that everyone knows about but nobody documented because it was embarrassing.
You hand them a sticky note that says "fix the login thing."
And then this is the part that gets me you are frustrated with THEM.
That is the average prompt. "Fix the login thing." No files. No context. No definition of done. No acknowledgment of Dave and his crimes against the session store.
Tokenmaxxing is writing the actual brief. Every time. Even when you are in a hurry (especially when you are in a hurry, honestly, but that is a different post).
The three things (I tried to make it not three things, it is three things)
Context architecture — what actually goes in the window. Not everything. The right stuff. There is a version of this where you dump your entire repo in and pray and there is a version where you load the two files that matter and explain why they matter and one of those versions works a lot better. (It is the second one.)
Prompt structure — the shape of the ask. Constraint up front. Format specified. One worked example if you have one. One job per prompt, not seven, please, for the love of — anyway. Specific in, specific out. Vague in, vague out. This is not complicated but it requires you to know what you actually want before you ask, which turns out to be the hard part.
Codebase legibility — okay this one people do not talk about enough. How readable is your project to something encountering it cold? Because if your files are named helpers2_new_ACTUALLY_FINAL.js and your functions are two hundred lines long and your folder structure is a historical artifact of decisions made in 2019 by someone who left — the model is going in blind. Clean structure was always good practice. Now it also directly affects your output quality. Two reasons to do the thing you should have been doing anyway!
The part that is going to annoy you
Tokenmaxxing is just clarity.
Specific inputs. Stated constraints. Examples. One job at a time. These are the same things that make a good ticket. A good PR. A good document that humans can actually read without needing a decoder ring.
The model did not invent this standard. It just enforces it immediately and without mercy instead of letting your vague thinking float around for three sprints until it explodes in production.
I think about this a lot. It keeps me up a little.
Go tokenmaxx something
The gap right now is not access to the tools. Everyone has the tools. The gap is in the inputs.
Two people. Same model. Same task. One of them thought about what they were putting in. One of them did not. You can tell which output is which by reading the first three lines.
You have the car. The car is fast. I am begging you to learn to drive the car.
Top comments (0)