@Mnilax - Boris Cherny Token Waste

Open original

@Mnilax on X The post promotes a podcast where Boris Cherny, creator of Claude Code at Anthropic, explains how common habits burn through Claude's token budget before the model even reads your prompt. The author adds their own breakdown from 400+ hours of usage and suggests heavy users likely hit several of these patterns.

  • Cherny reportedly attributes about 14% of token loss to CLAUDE.md files loaded before any typing happens.
  • Re-reading old chat history accounts for roughly 13% of wasted tokens, and hooks you forgot you installed add about 11%.
  • The post argues that most "Claude got dumber" complaints are misdiagnosed rather than real model regressions.

The post is a promo for a podcast in which Boris Cherny, who created Claude Code at Anthropic, supposedly breaks down nine patterns that waste about 73% of your tokens. The claims as stated: roughly 14% of tokens go to CLAUDE.md before you type anything, about 13% goes to re-reading old chat history, and about 11% comes from hooks you installed and forgot about. The author also says that most "Claude got dumber" complaints are misdiagnosed, and that heavy users hitting Max plan limits more than once a week probably have four or more of these patterns, maybe seven. None of the figures are sourced from anything beyond the post itself, so treat the percentages as the author's framing rather than measured results. The article is mostly a hook for the video, and the substance is thin in the post. Still, the underlying idea is worth taking seriously: a lot of token spend happens before your prompt is ever processed. Instruction files, accumulated conversation context, and background hooks all get sent along with each request, so bloat in any of them compounds across a session. The practical takeaway, if the numbers hold up, is to audit what's being loaded each time: trim CLAUDE.md to what you actually need, start fresh sessions instead of dragging long threads forward, and check which hooks are running. The author says they've spent 400+ hours on this and will share a breakdown after the podcast, so the detailed evidence lives elsewhere. Worth saving as a lead to check, not as a settled guide. The specific percentages and the "73%" headline are unverified, and the link is a promotional post designed to drive views. If you want to act on it, pull up your own usage stats and see whether the context-loading overhead is as large as claimed before changing anything.

Ask about this article