Getting the most out of Opus 5.5 in Claude and Claude Code

Practical instructions for working with Opus 5.5 in Claude and Claude Code, emphasizing that the model runs longer, explains its actions, and automatically “thinks” before replying. Key prompting rules: give the whole task in one message, define a concrete finish line (e.g., tests pass), and specify when to stop or ask. Remove “think carefully” style prompts because Opus 5.5 internally manages deliberation and starts replies sooner without them. For design requests, list specific styles to avoid. Allow mid-run additions without restarting, and prefer explicit finished-file requests (spreadsheets, documents) rather than outlines. Early testing shows major gains on multi-step work and better bug-catching at lower effort than prior models.

Operational specifics for long runs and verification: add stop/continue rules to CLAUDE.md so the model only pauses for needed input or before destructive actions; have Opus split large audits across subagents and require each to provide evidence; keep a running TASKS.md checklist so progress survives context summarization. When a run finishes, read what Opus is waiting on first, then its summary; ask it to mark anything it couldn’t confirm and to review diffs or PRs explicitly (file, line, why, how to reproduce). In Claude apps, attach charts or screenshots for accurate visual interpretation and instruct the model to treat settled answers as final to avoid backtracking. Note: Opus 5.5 launches with stricter bio/cyber safeguards that can reroute flagged messages to older models; routine security and health queries remain supported, with tuning ongoing.