Prompting Claude Opus 5.5
Anthropic's official guide outlines prompting patterns, effort calibration, and harness adjustments for Claude Opus 5.5 to manage behavioral changes and prevent early stops.
summary is enough Developers integrating Claude Opus 5.5 can quickly grasp the key prompt adjustments and harness changes from the summary without reading the full documentation.
summary
Anthropic's guide outlines how Claude Opus 5.5 differs from Opus 5 and provides specific prompting and harness patterns to address them. Key areas include calibrating the always-on thinking effort, handling unattended agent tasks to prevent premature stops, managing new safeguard refusals, and marking pasted text to resist prompt injection. The document serves as a migration manual for developers encountering unexpected model behaviors or high token consumption.
key points
- Effort level replaces manual thinking toggles, with medium matching previous high performance at lower cost and latency.
- Unattended agents require explicit checklist tracking and system prompts to prevent text-only turns from halting execution prematurely.
- Pasted user text must be wrapped in randomized tags with specific system instructions to defend against indirect prompt injection.
- New safety classifiers and reasoning extraction restrictions require reading thinking blocks rather than prompting the model to output internal reasoning.
who should read it
Software engineers and AI application developers migrating systems to Claude Opus 5.5.
discussion skeptical
consensus
Commenters express fatigue over constantly changing prompting patterns and shifting model behaviors, with many considering migration to alternative providers like DeepSeek due to strict safety refusals and usage limits.
pushback
A minority find the rapid iteration exciting, arguing that model-mediated automation and advanced harnesses are worth the engineering overhead.
All that keeps jumping out at me is how they've set it to refuse giving users thinking tokens and prompts for full reasoning in output. Just drives me further away; I may not stop using Claude completely for now, but I'll be moving even more of my primary workload to Chinese providers. That's where openness and freedom is now at.
Highlights developer frustration with safety refusals and reasoning restrictions, pointing to a broader trend of moving workloads to alternative providers.
Generated by an LLM from the article and thread. Check the sources before relying on it.