Anthropic Publishes Official Prompting Guide for Claude Opus 5.5: Calibrate Effort, Don't Max It
CodeByPoonam · x · 2026-09-25
Anthropic released an official prompting guide for Claude Opus 5.5 in its platform docs. Key points: output is 30%+ faster than Opus 5 with fewer tokens per task; start at medium effort—it already beats Opus 5's high at lower cost; thinking is always on so 'think carefully' is obsolete; unattended agents report progress mid-task and need explicit 'keep going'; mark pasted content separately from instructions; ban specific frontend patterns instead of vague 'make it unique'; the model now reads charts and screenshots accurately; the guide also covers safeguard refusals, multi-agent time signals, and multi-app workflows.
More from coding & agent
- Beyond embeddings: grep agents and whole-corpus long-context prompts as underrated retrieval patterns — jakedahn · 2026-09-25
- Nokia Open-Sources AnyJev: Turn Any LLM into a Calibrated Decision Model, No Training — kalyan_kpl · 2026-09-25
- Microsoft's biggest Copilot update ever: Autopilot agents, Code, Home, and embedded Office — satyanadella · 2026-09-25
- TSP: Self-Hosted LLM-Powered A-Share Quant Workbench Passes 5,000 GitHub Stars — shy3130 · 2026-09-25
- Starnet: Pixel-Art Desktop Harness for Watching AI Agents Work, Built Local-First — androoAGI · 2026-09-25
- huashu-chrome: open-source MCP server lets AI agents drive your logged-in Chrome with 22 tools — AlchainHust · 2026-09-25