Boris Cherny, the creator of Claude Code at Anthropic, just listed 9 patterns that waste 73% of your tokens. in this podcast he breaks down exactly how the model burns tokens before it even reads your prompt: - the 14% you lose to CLAUDE.md before typing a word -... the 13% you pay re-reading old chat history - the 11% from hooks you forgot you installed - why most "Claude got dumber" complaints are wrong if you're hitting Max limits more than once a week, you have at least 4 of these. Probably 7. instead of another show tonight, watch this. my own breakdown based on 400+ hours of usage is below, read it after the podcastshow more

I figured this out a while back, so I built Token Optimizer. It audits & fixes your structural + runtime waste, tracks *context quality*, has smart compaction + session continuity (so you don't lose anything important), and even has coach mode to help you improve. + You get a dashboard with full visibility into your sessions & waste patterns. Works with OpenClaw + Codex too:

@RMB The 50k tokens you spend on system prompt and tools every time you boot 4.7

Stole this content from @lennysan without any credit at least give him a shoutout if you’re going to use his video

Here's the thing though. The "you are using it wrong" excuse needs to eventually be solved in a way that a user doesn't have to think about how they are using it and can just use it. I get it has a lot to do with hardware limitations but it eventually needs to be solved if AI is going to really become a major part of all our lives. We basically need to have that jump from the days of dial internet and keeping track of AoL minutes to always online high-speed cable internet.

I am the user so, I am not changing my habits because someone "Suggested" I change my habits to however they want me to use them. Antropic is wrong, you make the product for the consumer, you cannot shape the consumer to your product. It is the other way around, and this misinformation warfare of how I should think, how I should behave to be a better user is BULLSHIT.

73% "in context" isn't waste. It's working memory. CLAUDE.md is a calibration tax that pays back in fewer correction tokens later. The real metric is signal density per turn, not % of non-prompt tokens. RAM isn't "wasted" on the OS.

Seems like a problem with the tool if it can't identify and handle these automatically? or possible they dont give a shit because people spend more money because of a shit tool

I stopped listening to him after “safety”. I want complete creative control. It’s like telling a standup comic to be politically correct.

You should be able to use the product without having to worry about token consumption. You don’t drive your car and constantly worry about acceleration, coasting and lifting to save fuel.

trimmed my CLAUDE.md from like 600 lines to 200 last week and stuff stopped feeling laggy, had no idea it was the bottleneck

so… continues to telling the users that im using it wrong and its a skill issue alright man

Yes, if that's what is happening, that MEANS claude got dumber. It could write hooks, install tools, use computer on demand, and it doesn't. Blame me all you want. It's dumb

So one of anthropics biggest products came from no roadmap or product scope?

Instead of making a long ass video why don't you adapt your product to not waste tokens like a bunch of idiots

basically: -CLAUDE.md bloat: 4800 tokens loaded every single turn. -Conversation re-reads: History re-tokenized on every follow-up message. -Hook injection waste: Plugins prepend 6200 tokens per prompt. -Cache misses: 5-min timeout forces full re-tokenization on resume. -Skill loading: Irrelevant skills load 13500 tokens just in case. -Tool definitions: 12 MCP schemas add 7200 tokens always. -Extended thinking: Unneeded reasoning burns 3000 tokens on simples. -Wrong-direction generation: Finish bad 400-line outputs wasting tokens. -Plugin updates: Auto-updates inject redundant context at session start.

Media trained to the max

Just use codex and forget about usage limits

Or move to Codex, use a more capable model without having those extreme rate limits

That’s the crux. If I have to think so much to use a LLM, then I would rather code. Make models raw and kill customizations like Anthropic did to destroy their best model.

Made a Kindle-ready transcript of this interview (Lenny's Podcast x Boris Cherny). Cleaned up, 26 section headers, clickable TOC. EPUB download: Built with an open-source Claude Code skill:

Run the audit script, then aim for: CLAUDE.md: under ~1,200 words Active plugins: 3–5 max Active skills: 3–5 max Always-on MCPs: around 3 Chats: compact or restart around 20 messages Basically: cut the context tax, and your same plan feels 2–3× bigger.

8 minutes in and this guy is acting like curso never existed

We can replace CLAUDE.md with Archcore. Multiple files instead of one, plus we load them at session startup, which is useful.

maybe instead of blaming the user they should try to improve the harness to solve how users wants to use it

Really enjoyed this podcast on scaling Claude workflows 🧠 @lennysan 
The best part was the 9 patterns that cut costs and stop bloated prompts/history bloat. Super practical for anyone in the agentic space. 
Full summary 👇

@threadreaderapp unroll

The 11% from hooks you forgot you installed is personally attacking me right now.

He was great in breaking bad

seems to me like a harness concern, not a user problem

Boris Cherny listing 9 patterns that waste 73%% of tokens is the Claude Code efficiency guide everyone needed. The 14%% burned on CLAUDE.md before you type a word is the silent killer. Context window budget is the new performance budget.

I felt that 14% CLAUDE. md punch personally 😂

This entire post is like saying I’m selling you a car. You’re probably driving it wrong and not addressing the fact that Anthropic never invested as much in servers as open AI or Google. That’s the reason they can’t handle that much from clients. That’s the reale issue.

@grok Analyse this videoand create me ten highly useful prompts I can use for Claude code.

Codex is the best