Anthropic Cut 80% of Claude's System Prompt Without Losing Performance
nptacek · x · 2026-08-10
Anthropic recently shared insights into context engineering for Claude, revealing that they deleted 80% of Claude Code's system prompt for newer models without a drop in performance.
Key shifts include abandoning rigid rules in favor of model judgment, stopping tool examples (which constrain the model), adopting progressive disclosure to load skills only when needed, and providing rich references like code and test suites instead of simple markdown specs.
More from coding & agent
- Developer Releases Background Agents: Run Coding Agents in Daytona Sandboxes — jamesmurdza · 2026-08-10
- Open-Sources Agent Workflow for Automated Claude & Codex Code Review — Ranteck · 2026-08-10
- Crime Economists Host Hackathon to Batch Generate Paper Drafts with AI — paulnovosad · 2026-08-10
- Claude Code output styles cut token usage by 50% while boosting explainability — daniel_mac8 · 2026-08-10
- AgentSkillOS: Build AI agents from 200,000+ skills via retrieval and orchestration — tom_doerr · 2026-08-10
- Using OpenAI Codex as an Agent: Automating Database Migration in an Hour — leebase65 · 2026-08-10