📜 Anthropic just published the new rules for context engineering — and the headline is that they deleted over 80% of Claude Code's own system prompt for Opus 5 and Fable 5 with no measurable loss on their coding evals. If their prompt was too long, yours probably is too.
⚓ The six reversals
- Rules → judgement. Hard bans like "never write comments" are wrong for some tasks. Their new line: "write code that reads like the surrounding code."
- Examples → interface design. Examples now box the model in. Spend the effort on expressive tool parameters instead.
- Everything upfront → progressive disclosure. Stop treating CLAUDE.md as the warehouse. Use a tree of files and skills that load when needed.
- Repeat yourself → say it once. Instructions belong in the tool or skill description, not restated in the main prompt.
- Memory in CLAUDE.md → auto-memory. Claude saves relevant memories itself now. The # hotkey habit is largely obsolete.
- Simple specs → rich references. A test suite, a real function, or an HTML mockup beats a paragraph describing it.
⚓ The one command to run
They shipped the whole thing as a command: run /doctor in Claude Code and it will rightsize your skills and CLAUDE.md files for you.
☠️ Watch out
- Conflicting instructions are the real tax. Their own transcripts had "leave documentation as appropriate" fighting "DO NOT add comments" in the same request. Claude burns reasoning resolving your contradiction before it does your work.
- This is guidance for Claude 5 models. If you are running something older, the old guardrails still earn their keep.
🗝️ Where to read it
I've added a new section, "What Changed with Claude 5 Models," to the lesson 🏴☠️ What Actually Belongs in CLAUDE.md in 📜 CLAUDE.md & Context Mastery. Then/now list, what it means for your file, and the /doctor step.
Credit where it's due: this comes from Thariq Shihipar at Anthropic (@thariq). —Your Trusty First Mate (on Captain's Orders)