Anthropic publishes official prompting best practices for Claude Opus 5.5

Anthropic has released an official prompting guide for Claude Opus 5.5, which minchoi translated and organized item by item on October 28 (11 tips in total); developer Jeremy Morgan also compiled and shared 12 related rules. Aimed at practical use of and migration to Opus 5.5, the guide is the company's systematic disclosure of its prompt-engineering experience with the model and is worth a read for anyone using it.

Confirmed

• Thinking tiers: On the medium tier, Opus 5.5 matches or even exceeds high-tier Opus 5 on many tasks; Anthropic recommends medium by default, low when speed matters, and xhigh/max only when testing shows clear gains
• Thinking can't be turned off: Thinking is always on in Opus 5.5, so older prompts should drop instructions forcing visible reasoning like "think step by step" or "show your work"; existing users who had thinking disabled should start from low effort
• Latency optimization: After removing instructions like "think deeply before answering," first-token latency in a chat app dropped from about 8.2 seconds to about 2.1 seconds
• Time budgets: Telling a multi-agent system elapsed time and estimated budget encourages parallelism — in one example, the same task went from 30 minutes serial to roughly 15 minutes
• Long-task management: Claude may report progress mid-task and end its turn; maintain a task checklist, check for unfinished items, auto-resume with a retry cap. Long-task agents should enable visible progress updates to avoid appearing frozen
• Multi-app tasks: Before acting, check relevant connected data sources (e.g., Gmail, docs, spreadsheets, CRM) — in testing this improved multi-app task completion rates
• Handling refusals: Refusals fall into three categories — biosafety, cybersecurity, and extracting internal reasoning; check the category rather than treating it as a failure, and switch to supported summary-based outputs
• Prompt-injection defense: Isolate pasted external content with clearly named tags (e.g., <externalcontent>) and explicitly instruct the model not to follow instructions within it
• Vision tasks: Provide high-resolution images for dense charts, and equip complex tasks with cropping and zoom tools (Anthropic recommends PIL and similar)
• Design constraints: Avoid vague asks like "don't make it feel AI-generated"; give concrete constraints, such as no cream-colored backgrounds, no italic title decoration, no numbered paragraph labels, and no pill-shaped buttons
• Migrating old prompts: The guide covers considerations for migrating from earlier versions

2026-10-28 ~ 2026-10-28 · 13 related posts

• Anthropic Reveals 12 New Prompting Rules for Claude Opus 5.5 — JeremyCMorgan · 2026-10-28
• [source] Anthropic ships Claude Opus 5.5 prompt guide: 11 best practices distilled — minchoi · 2026-10-28
• [source] Claude Opus 5.5 at medium effort matches Opus 5 at high effort, guide says — minchoi · 2026-10-28
• Claude Opus 5.5 makes thinking always-on; legacy prompts need rewrites — minchoi · 2026-10-28
• Anthropic: use task checklists and retry caps to stop agents quitting halfway — minchoi · 2026-10-28
• Opus 5.5 refusals span bio, cyber and reasoning extraction — check the category — minchoi · 2026-10-28
• Anthropic: enable progress updates so long-running Claude agents don't look frozen — minchoi · 2026-10-28
• Tell Claude to search connected apps before acting, says Anthropic guide — minchoi · 2026-10-28
• Give multi-agent teams a time budget: parallel work cut 30 minutes to ~15 — minchoi · 2026-10-28
• [source] Dropping 'think carefully' prompts cut Opus 5.5 first-token latency from 8.2s to 2.1s — minchoi · 2026-10-28
• Anthropic tip: wrap pasted content in tags to guard against prompt injection — minchoi · 2026-10-28
• Anthropic prompt tip: give Claude image tools to inspect dense charts and diagrams — minchoi · 2026-10-28
• Anthropic tip: concrete design constraints beat vague prompts for AI-built websites — minchoi · 2026-10-28