More instructions were not making Claude better
Anthropic cut more than 80% of Claude Code’s system prompt for the Claude 5 generation models, with no drop in coding evals.
What interests me isn’t the number. It’s the admission behind it: they were overconstraining the model, through the system prompt, CLAUDE.md files and skills alike, and the competing instructions made it work harder, not better. That is exactly what I keep relearning writing CLAUDE.md files and skills for Mate and for my own AI assistant. Fewer explicit rules, richer references over descriptions, context loaded when it’s needed instead of all of it up front.
On https://claude.com/blog/the-new-rules-of-context-engineering-for-claude-5-generation-models
// MORE SHORTS
Somebody still has to read it →
A post proposing AI;DR hit #1 on Hacker News: if you didn't read and edit what the model produced, don't expect the author to read it either. Producing a wall of text became free.
// MORE SHORTS
How I Made Symfony's Maker Agentic →
Twelve merged pull requests in one release, my first-ever contribution to Symfony's maker-bundle. All of them were about the same thing: agents kept running into make:* commands that only worked with a human answering prompts, and simply crashed otherwise.