Prompting Claude Opus 5.5
5 hours ago
- Claude Opus 5.5 generates output tokens more than 30 percent faster than Claude Opus 5 and tends to finish the same task with fewer tokens.
- Effort is the main control for thinking, starting at medium (default), with low and max levels available for tuning latency/cost vs. quality; lowering effort reduces thinking more reliably than prompt instructions.
- Claude Opus 5.5 does not accept thinking: {"type": "disabled"}; prompts written for thinking-disabled integrations need adjustments like starting at low effort, removing reasoning-extraction instructions, and reading thinking blocks with display: "summarized".
- For unattended agentic runs, treat text-only end-of-turn messages as progress reports, keep a checklist, and add a system prompt instruction preventing premature stops like summaries or offers to wait.
- Safeguard refusals include biology, cybersecurity, and reasoning_extraction categories; requests asking the model to reproduce internal reasoning in response text can be declined.
- Progress updates between tool calls come as thinking blocks with empty text by default; set display: "updates" to receive summaries, give the model a tool for verbatim content, and add reminders if turns go silent.
- In multi-app workflows, add a system prompt instruction to explore broadly (list and open emails, documents, spreadsheet tabs, records) before acting, which improves task completion.
- Use time signals in multiagent harnesses—give elapsed time or a budget (e.g., 'elapsed 340s / 1200s')—to encourage parallel work and faster completion.
- In chat system prompts, remove 'think carefully before answering' lines to reduce latency, and optionally add an instruction to treat earlier answers as settled to avoid re-thinking on follow-ups.
- Mark pasted text in user messages with <pasted_content id="..."> tags and a system prompt note to resist indirect prompt injection from pasted instructions.
- Re-test visual scaffolding: Claude Opus 5.5 reads charts/diagrams/screenshots more accurately without tools; for dense inputs, use higher-resolution images and image-processing tools like cropping with PIL/OpenCV.
- For frontend design, name specific styles to avoid (e.g., cream backgrounds, italic accent words, numbered labels, monospace labels, pill buttons) instead of generic 'avoid AI look' instructions.