Anthropic cut 80% of Claude Code's system prompt for Claude 5 models with no coding eval loss
cfregly · x · 2026-11-01
Anthropic engineer Thariq Shihipar details new context-engineering rules for Claude 5-generation models: the team removed over 80% of Claude Code's system prompt for models like Claude Opus 5 and Claude Fable 5 with no measurable loss on coding evals — old prompts were overconstraining newer models.
Key points:
- Context engineering spans system prompts, Skills, CLAUDE.md, and memory — it's shared across requests, so it can't be as specific as a prompt.
- Tool-use examples can limit exploration in newer models; keep expressive tool parameters and only retain examples that testing proves useful.
- A new claude doctor command (/doctor in Claude Code) rightsizes your skills and CLAUDE.md files.
- Anthropic's updated prompting best practices docs now include per-model guidance (effort levels, tool-call batching in agent loops, long-run progress reporting).
More from coding & agent
- Anthropic publishes Skill authoring best practices: keep SKILL.md under 500 lines — bibryam · 2026-11-01
- Dev builds entire Three.js city with no 3D models using Codex and GPT-6 Astra — OpenAIDevs · 2026-11-01
- Compiler-backed coding agent Benzi hits 82.6% on SWE-bench Verified for $9.26 — DonkeyTheKing · 2026-11-01
- Ronie: free open-source AI agent that makes Ray-Ban Meta glasses actually smart — Odd-Sun6296 · 2026-11-01
- GPT-6.1 Terra builds 3-level game in 6.5 min; DeepSeek V4.1 Flash slower but 36% cheaper — latent_lurker8587 · 2026-11-01
- Theo: switching Opus from xhigh to high reasoning actually works better — andersonbcdefg · 2026-11-01