Anthropic rewrites its context engineering rules for Claude 5
In a new blog post, Anthropic’s Tariq Shihipar lays out how context prep has shifted for the Claude 5 generation and which older recommendations no longer hold. The new guidance:
• Trust the model’s judgment over rigid rules. An early Claude Code system prompt banned multi-line comments; the instruction now just asks it to match the surrounding style. • Design a tool’s interface and expressive parameters instead of feeding it usage examples. • Use progressive disclosure: some tools load only after a ToolSearch query. • Auto-memory, which Claude saves on its own, replaces the CLAUDE.md memory file.
Anthropic also shipped a claude doctor command to clean up prompts, skills, and CLAUDE.md files.
Up next
Opus 5 tops Fable 5 on OSWorld 2.0 at a third of the price
Opus 5 beats Fable 5’s best OSWorld 2.0 computer-use score at a little over a third of the cost: $5 per million input tokens, $25 per million output. On ARC-AGI 3 it scores triple the nearest model.
Related stories
- Claude Code optimization: choosing models and effort levels
- Claude Code has four kinds of agent loops: /goal, /loop, /schedule
- Claude Code cut over 80% of its system prompt, scores held
- Delete “think carefully” from your Opus 5.5 prompts
- “Verify twice” wastes tokens on Opus 5, Anthropic says
- A hidden system prompt paragraph tells Claude Code to stop asking permission
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.