AI Pulse · Under the Radar

Anthropic just proved you can throw away most of your system prompt and Claude still works fine. They cut more than 80% of the instructions they were feeding Claude Code for Opus 5 and Fable 5, and performance held steady. Developers noticed, the story hit Hacker News twice, and Anthropic confirmed it on their own blog.

The new playbook is simple. Stop cramming detailed instructions and examples into a static system prompt or a CLAUDE.md file that sits there all the time. Instead, build better tools and load skills or context only when the model actually needs them. If Claude needs to know how to format a table, give it that instruction when a table shows up, not upfront in a wall of text it may never use.

This matters because most teams are still writing prompts like instruction manuals. They front load everything, hoping the model remembers it all. That burns tokens, slows responses, and often confuses the model with irrelevant detail. The smarter move is a thin router that pulls in the right pipeline or tool at the right moment. If you are running agents or automation at any scale, this is the kind of architecture we build into an AI command centre, where context gets managed dynamically, not dumped upfront.

If you have a CLAUDE.md or a long system prompt in production right now, it is worth auditing. Ask what the model actually uses versus what you are paying to send every time. Odds are you can cut half of it, route the rest on demand, and get the same or better results. Anthropic just showed that less instruction, better tooling, and smarter context loading is the path forward.