Claude Opus 5.5 Prompting Guide

Produce a usable prompt or instruction change for Claude Opus 5.5. Prompts written for Opus 5 carry over and remain the baseline; this package adds what changed in 5.5 and removes what 5.5 no longer needs. It applies Anthropic's Prompting Claude Opus 5.5 together with Prompting Claude Opus 5, which the 5.5 page names as its starting point. Loading it does not select a model, set effort, or change any runtime.

Put Each Fix In Its Layer

Much of the Opus 5.5 guidance changes a setting, the agent loop, or a tool rather than prompt text. Decide the layer before writing a clause:

• Prompt text: scope, completion, stop policy, communication cadence, exploration, and the instructions to remove.
• Request settings: effort, max_tokens, thinking.display, per-message effort, and refusal fallback. Prose cannot stand in for them; a line asking for less thinking works less reliably than lowering effort.
• Agent loop: what counts as the end of a task, continuation messages, silence reminders, time signals, and how instructions or tools change without editing earlier turns.
• Tools: a send-to-user tool for exact mid-turn content, cropping or a container for dense images, and strict tool use for schema-valid calls.

Then identify who owns each layer in the target host. An API integration owns all four. Instruction files loaded by Claude Code or the Agent SDK own the text and a few settings; the harness already runs the loop, keeps history append-only, and marks pasted text, so restating those in CLAUDE.md or a skill adds context without changing behavior. A chat product owns its system prompt and request settings. Read host notes before editing instructions that Claude Code or another agent harness will load.

Remove Before Adding

Most Opus 5.5 guidance deletes. Before adding a clause, look for these in the existing stack and remove or re-test them: