its crazy to me that people are not even running editorial passes. its trivial to give claude a 5 line instruction about writing style and have it transform the output into something vastly better.
All of these token optimiser things...i rarely see any performance analysis. If you're cutting token usage in half, what does that do to the accuracy of the outcome? If you're automatically delegating to cheaper models, what does that do to the outcome?
Personally I've found a HUGE improvement in quality using more context, to the point that it's 'better value' - fewer outcomes I disagree with or have to fix/rework.
I've mostly done this through mechanical documentation links that are enforced and can be 'followed' with some python files that the agents use. Context is higher than just reading code, but it has been a HUGE benefit (as far as I can tell).
Just use billion context. I customize it to set a slider so I decide how big the context window can ever grow, and it uses models to fold itself in on itself and summarize itself while retaining the entire chat history. I've managed to run chats over three thousand messages long without any negative effect. And overall, I average a seventy-four percent reduction in token usage.
Please don't let AI write your README.md. It's illegible to humans.
its crazy to me that people are not even running editorial passes. its trivial to give claude a 5 line instruction about writing style and have it transform the output into something vastly better.
All of these token optimiser things...i rarely see any performance analysis. If you're cutting token usage in half, what does that do to the accuracy of the outcome? If you're automatically delegating to cheaper models, what does that do to the outcome?
Personally I've found a HUGE improvement in quality using more context, to the point that it's 'better value' - fewer outcomes I disagree with or have to fix/rework.
I've mostly done this through mechanical documentation links that are enforced and can be 'followed' with some python files that the agents use. Context is higher than just reading code, but it has been a HUGE benefit (as far as I can tell).
I use this and it works well: https://github.com/rtk-ai/rtk
I'm trying to figure out if I can use multiple of these tools together without conflicts
JetBrains benchmarked RTK in July and found that it doesn't work;[0]
"rtk advertised saving: 60–90%. Measured on real agent work: +7.6% more expensive at low reasoning effort (p=0.004), ±0% at high effort."
[0] https://blog.jetbrains.com/ai/2026/07/rtk-claude-code-token-...
Can you define "well" at all? I personally think a lot of these are absolute snake oil.
People throw them in and never think about it. Sure it doesn't BREAK things, but it could degrade them by 15% and you'd never know?
Just use billion context. I customize it to set a slider so I decide how big the context window can ever grow, and it uses models to fold itself in on itself and summarize itself while retaining the entire chat history. I've managed to run chats over three thousand messages long without any negative effect. And overall, I average a seventy-four percent reduction in token usage.
> https://github.com/localixai/tokendoctor/tree/main/bench
Don't we need to run it on all of the benches those models publish?
This doesn't look very comprehensive. Moreover half of the times factory claude code is better than 'advised' one.
Although I don't discourage such experiments, I feel this one is way complicated for very little or no gain provably.
I have a list of ai related tools. There is so many of them
For example limiting token usage i think can also be achieved by 'caveman' project. It achieves it differently but the goal is same i suppose
Overall Ai tools I found:
https://github.com/rumca-js/awesome-database-top/blob/main/l...
well then don't go https://neuronary.org/ or https://www.neuronpedia.org/ you'll never be done.
These tricks last at most one model dot release.
Also not measuring cost per task is one of the sins. It's easy to get penny wise and pound foolish on this matter.
why not having a tiny fast model summarize tool output instead, maybe with a threshold or something.