AI Summary
This video addresses the common problem of rapidly exhausting Claude Code usage limits. The creator identifies four hidden 'leaks' that drain tokens—tool output, verbose replies, overbuilt work, and sending small tasks to the most expensive model—and presents four free, open-source tools to plug each leak, claiming users can reduce token usage by up to 80%.
Chapters
Most users respond to usage limits by typing shorter prompts or asking fewer questions, but this is the wrong fix. The real drain comes from four hidden leaks in how Claude Code operates.
When Claude runs commands, the raw output floods back into the chat, including giant logs and walls of text. This entire context is carried along with every new message, consuming tokens.
Claude tends to be overly polite and verbose, writing multiple paragraphs when a few words would suffice. This fluff consumes usage tokens.
Claude often builds extra features or structures that weren't requested, 'just in case.' This extra code is then read back later, wasting tokens.
Users often send trivial tasks (renaming, tiny summaries) to the most powerful model when a free or smaller model could handle them, wasting expensive tokens.
A free, open-source tool that acts as a filter between Claude Code and your computer. It rewrites commands, trims output, and collapses repeated lines. The team claims it can cut up to 90% of command output, and the creator's test showed an 82.9% reduction in tool output tokens.
A free skill that makes Claude's replies short and direct, like a caveman. It has levels from light to ultra. The creator's test on Fable 5 showed 69% fewer output tokens, cutting overall token usage by about 30%.
A free skill that prevents overbuilding by making Claude act like a lazy senior developer—efficient, not careless. It reads code first, then builds only what was asked. Project testing showed it wrote about 54% less code and ran roughly 20% lighter.
A free, open-source gateway that runs locally. It routes each request to the most appropriate model, sending grunt work to free models and hard work to Claude. It connects to hundreds of providers, with over 90 offering free tiers. It also includes built-in compression.
Mistake 1: Rationing prompts (wrong fix). Mistake 2: Expecting 80% savings on every task (varies by workload). Mistake 3: Installing the wrong RTK (there are two projects with that name).
Tip 1: Type '/clear' after finishing a task to wipe the conversation and start fresh. Tip 2: Trim your Claude.md file, as every rule in it is read at the start of every session.
By identifying and plugging the four main token leaks with free tools, users can significantly extend their Claude Code usage. The key is to fix the plumbing first, not ration your prompts, and to be mindful of common setup mistakes.
Mentioned in this Video
Tutorial Checklist
💡 Key Takeaways
The Real Leaks Are Hidden
Shifts the focus from user behavior to systemic inefficiencies in how Claude Code operates, providing a more effective solution.
00:57RTK's Claimed 90% Reduction
Provides a concrete, verifiable claim that can be tested by users, adding credibility to the tool's effectiveness.
02:10Ponytail's 'Lazy Senior Developer' Approach
Uses a memorable analogy to explain a technique that reduces code bloat while maintaining quality.
04:47Omniroute's Model Routing
Introduces a smart routing strategy that leverages free models for simple tasks, a scalable approach to cost optimization.
05:30The '/clear' Command Tip
A simple, actionable tip that can immediately reduce token usage by preventing context bloat.
07:42Full Transcript
[00:02] 80% for free. Who's actually reading all that output in your Claude code session? Not you, Claude is. Every log, every wall of text, every polite paragraph, and each line eats into your usage. Four free
[00:16] tools stop it, and they all run inside my agent OS. One of them even hands the boring work to free models. In this video, I'll show you the four hidden leaks draining your Claude code usage. Then I'll show you the four free tools I
[00:30] use inside my agent OS to plug everyone. Stick around because the last tool is the one that surprised me most. And near the end, I'll share the mistakes that undo all of this, plus two quick tricks you can use today. Here's what most
[00:43] people do when they hit their Claude code limit. They type shorter prompts, they ask fewer questions, they ration like it's water in a desert. That's the wrong fix. The words you type are a tiny slice of the problem. The real problem
[00:57] is four hidden leaks. Leak one is tool output. When Claude runs a command on your machine, the raw result floods back into the chat. Giant logs, walls of text Claude never needed to read. Every line counts, and it doesn't just sit there.
[01:13] Everything in the chat gets carried along with every new message. Leak two is Claude's replies. Claude is polite, really polite. It will write three paragraphs when three words would do. All that fluff comes out of your
[01:26] usage, too. Leak three is overbuilt work. Ask for one simple thing, and Claude often builds five extra things just in case. Extra structure, extra options nobody asked for, and all of it gets read back
[01:39] later. Leak four is the biggest one. You send every small job to the biggest model. Renaming things, tiny summaries, little fixes. A free model could handle those just fine. Four leaks, four free tools, one tool for each leak. Plug all
[01:55] four, and the same usage goes a lot further. And I run all four my agent OS, so they work quietly in the background. Now, you don't need to be technical for this. Each tool installs with about one command. I'm not a coder either, so
[02:10] we'll plug them one at a time. Okay, plug number one, RTK. RTK stands for Rust Token Killer. It's free and open source. It sits between Claude code computer. Think of it like a filter on a tap. The water still comes through, but
[02:26] the junk doesn't. It's one tiny program, and it handles the commands Claude runs most, like Git, tests, and file listings. When Claude runs a command, RTK rewrites it, trims the output, and collapses repeated lines. Claude only
[02:42] sees the useful part. It works through a hook, so you set it up once, and it just runs. The team behind it says it can cut up to 90% of the command output your agent reads, and it adds under 10 milliseconds, so you'll never feel it.
[02:56] Quick pause before I show you the next part. If you want my setup, my agent OS, it's inside the AI Profit boardroom. The full agent OS zip file is in there, ready to install with this whole token saving stack already built in. We've
[03:10] also built a complete 30-day roadmap to go with it, so you can learn it step by step, and there are four coaching calls every week. If one of these tools doesn't click, you can bring your Claude code setup to a
[03:22] live call and get it sorted. Link in the description or go to ai
[03:37] It also works with other coding agents like Codex, Gemini CLI, and Cursor. I tested it myself. On my test, RTK cut the tool output tokens by 82.9%.
[03:49] The quality of the results stayed the same, so the team's claim held up almost exactly. Plug number two, Caveman. Caveman is a free skill for Claude code. It also works with lots of other coding agents.
[04:02] It makes Claude talk like a caveman. Short, direct, no fluff. You turn it on by typing slash caveman, and you turn it off by saying normal mode. There are levels, too, from light to ultra, so you choose how short you want it.
[04:17] Code stays exact, only the fluff goes. Here's an example from the project. A normal reply used 69 tokens. The caveman version used 19. Same answer. I caveman version used 19. Same answer. I tested it on Fable 5. I saw 69% fewer
[04:32] output tokens. That cut my overall token usage by about 30%, and every answer was still correct. One honest note. Caveman only shrinks what Claude writes. It doesn't touch the thinking, and the skill adds a small amount of its own
[04:47] each turn. That's why my overall number was lower than the reply number. Plug number 3. Ponytail. Ponytail is a free skill that stops Claude from overbuilding. It makes Claude act like a lazy senior developer.
[05:01] Not careless, efficient. Someone who's done this a thousand times and knows the shortest safe path. It reads the code first, then builds what you asked for In the project's own testing on real Claude code sessions, it wrote about 54%
[05:16] Claude code sessions, it wrote about 54% less code and ran roughly 20% lighter. And it keeps the safety basics. Error handling, validation, security, and accessibility stay in. It only skips the maybe we'll need this someday extras.
[05:30] Less code also means less for Claude to read back later, so every future session gets lighter, too. Now plug number 4. Omniroute. This is the big one. Omniroute is a free open source gateway that runs on your own machine. You point
[05:44] your coding tool at one local address, and Omniroute sends each request to the right model. It works with Claude code, Codex, Cursor, Cline, and more. It connects hundreds of providers, and more than 90 of them have free tiers. So
[05:59] grunt work goes to a free model. Hard work stays with Claude. wouldn' send your best person to sort the mail. If one free provider runs dry, omni route moves to the next one on its own. An RTK and caveman-style
[06:14] compression are built right into the gateway. That trims eligible requests by 15 to 95 %, which also helps with free model limits. One catch. If every task you run truly needs the top model, omni route helps you less. But be honest, how
[06:31] much of what you send Claude really needs Claude? You could set up all four by hand. Each one takes about one command and the whole stack takes around 15 minutes. But I run all four inside my agent OS. That way, every agent in there
[06:44] gets the compression, not just one. Claude, Hermes, and my other agents all work through the same setup, and it has an obsidian memory, so my agents already know my context. No re-explaining who I am at the start of every session. Fix it
[06:59] once and you never touch it again. Now, the mistakes I see people make. Mistake one is rationing. Shorter prompts, fewer questions. Your typing was never the leak. Fix the plumbing first and you get to do more with AI,
[07:14] first and you get to do more with AI, not less. Mistake two is expecting 80 % on every single task. It depends on the work. Tool-heavy sessions with lots of commands save the most because that's where RTK shines.
[07:27] My numbers came from my own tests, so yours will vary. Mistake three is installing the wrong RTK. There are two projects with that name. If RTK gain right one. Now, two quick tips. Tip one.
[07:42] Type / clear when you finish a task. It wipes the old conversation so your next task starts fresh instead of dragging dead weight along. Tip two. Put your Claude.md file on a diet. That's the instruction file Claude reads at the
[07:57] start of every session. Every rule in it gets read every time, so trim what you Here's something to think about. Right now, there's a window where most business owners haven't figured this out yet, and that window won't stay open
[08:10] forever. If you want to get ahead on learning this, the AI profit boardroom is built for it. Inside, you get the full agent OS zip file ready to install and the 30-day roadmap. You get daily tutorials so you keep learning as things
[08:24] change. Four coaching calls every week where you can ask about your setup live, a prompt library so you don't start from a blank page, and the member map so you can see the community you're learning with. Link in the description or go to
[08:38] aiprofitboardroom.com. If you want the full process, the SOPs, If you want the full process, the SOPs, and over 100 AI use cases like this one, join the AI Success Lab. It's our free community. Links in the comments and
[08:50] description. You'll get all the video notes from this episode plus access to 95,400 members. The AI Success Lab is free to join. The video notes pull everything from this episode into one place so you
[09:04] can follow along at your own pace. It's a simple place to learn, ask questions, and share what's working with other people learning AI. Join the AI Success people learning AI. Join the AI Success Lab and I'll see you in the next video.