Prompting Claude Opus 5.5

This is a practical prompting guide for Claude Opus 5.5 that describes how its behavior differs from Claude Opus 5 and recommends concrete changes to prompts and harnesses. Key points: Opus 5.5 generates tokens >30% faster and usually uses fewer tokens; thinking is always on and the default effort is medium (Opus 5 defaulted to high); the effort setting controls latency, cost, and how much the model deliberates, and should be calibrated against your own evals. Important operational details: thinking tokens count toward max_tokens (Anthropic suggests up to 128,000 for long agentic runs), use lower effort to reduce thinking, and use per-message effort changes to preserve the prompt cache. The guide also summarizes capability gains - stronger multi-step coding and code review, more accurate knowledge work and figures, improved reading of dense charts/diagrams and screenshots, and more reliable multi-step computer use.

The guide gives specific prompting and harness patterns: if migrating from thinking-disabled runs, start at low effort, remove instructions that force the model to reproduce internal reasoning, and read responses by block type. For unattended agentic tasks, treat text-only end_of_turns as progress reports (not task completion), keep a persistent checklist or to-do tool the model updates, send short follow-ups when items remain, and wait for background subagent outputs. It recommends system-prompt lines to reduce premature stops and setting display:"updates" to receive user-facing progress summaries. Other sections cover safeguard refusals, multi-app workflows, frontend defaults, marking pasted text, and tools for complex visual inputs; the migration guide lists breaking API changes.