What's this about?
Anthropic published a genuinely educational explanation of Claude Code's token economics — why output costs about five times input, and what users can do about consumption: run `/clear` between tasks so irrelevant prior context isn't resent, and set your model and effort level before starting, because changing either mid-conversation invalidates the cache. HN found the content useful and the framing insulting.
What exactly are prefill and decode?
A request goes through the GPU in two phases with very different costs. Prefill is the model reading everything you've given it — system prompt, CLAUDE.md, your message, files it read, command output. That happens in parallel and is comparatively cheap. Decode is the model writing, one token at a time; a 200-token answer is 200 sequential runs of the model. Each output token occupies the GPU far longer than each input token, which is why output is priced around 5x input. The practical implication: a verbose answer costs you more than a large input.
What's the community saying?
- "The Anthropic version of 'you're holding it wrong.'" The top-voted framing, referencing Apple's iPhone 4 antenna response — advice about user technique in place of a product fix.
- "Why not make a better harness by default?" If `/clear` between tasks and locking effort up front are the right behaviors, commenters asked why the tool doesn't do them.
- Unexplained cache rewrites are costing real money. A detailed report: at 400K tokens in, with a 1-hour cache TTL and no model or effort changes, cache writes jumped from small to 800K between messages — suggesting the cache invalidates for reasons users can't see or control.
- The enterprise-blame reading. A cynical but well-received interpretation: this becomes the document managers point at when the bill arrives, making token spend the engineer's fault.
- A real question nobody answered. Why is the prefix cache tied to effort level? A user running Fable at high effort for statistical modeling, then asking follow-up questions to understand the output, pays a cache penalty for a workflow that seems entirely reasonable.