Cut Your Claude Code Token Usage by 90% He leaked the exact methods. Most people will scroll past it. Nothing complicated. Bookmark this so you don't lose it.

Workflow beats brute-force spending.

This is pretty stupid, I just use claude code, and make judgement calls myself claude creates great solutions out of the box, and reviews it The judgement calls prompts he mentions, I just do myself, I don't need ai to tell me what direction we are going

The 90% claim tracks if you're currently pasting entire codebases into context. Most people are brute-forcing context when they should be surgical about it. The real unlock is treating context window like RAM — you wouldn't load your whole hard drive into a single program.

the token burn a big consideration in agentic coding right now. the people optimizing context and caching are quietly running the same workflows for a tenth of the cost, that gap compounds into an unfair advantage fast.

deepseek for code, Claude Opus for code review.

Kimi and Claude and Manus once in a while I use them for different tasks and outputs It was hard picking between Kimi and Claude but claude has more internal structure and the project option is a banger Real fire combo 🔥

I use rtk & caveman daily. I do have to optimize my rtk settings though since sometimes it’ll mangle the output so Opus is forced to bypass rtk with rtk proxy. Also started using That + caveman has been very interesting.

@kirillk_web3 wait, really? cutting by 90% sounds insane. gonna try this out asap. thanks for sharing!

looks useful

Model routing and being model agnostic is the future

I created a workflow management system to align goals with the agent, create gameplans, cascade changes through dependencies, and embed indexing making documentation the source of truth. It's had a similar effect while increasing accuracy.

Does this also work if I use an Opus 4.8 agent in the Cursor?

Can you give me a prompt to ask Claude Code to implement this and give me a dashboard to tune / edit when required?

@grok share the same format for codex, would like to use with gemini flash lite or deepseek flash, gpt got 5.5 high.

"The cheapest token is the one you never have to generate."

Muito bom, obrigado por compartilhar. Se conseguir fazer um sobre loop e goal seria ótimo também, pois são informações que se complementam.

worth reading

most devs watch the tutorial then ship with the same three commands

Most people don’t need more hacks, they need better discipline in how they use tools.

The real savings come from working smarter, not harder, sounds like a game-changer worth exploring.