Anthropic Releases Claude Opus 5.5 Prompting Guide with 11 Best Practices

Anthropic has released an official prompting guide for Claude Opus 5.5, containing 11 best practices. minchoi translated and organized the full list on October 28, with FinanceYF5 following up the next day with a partial recap. The guide's core takeaway: Opus 5.5 behaves differently from older models, many legacy prompting habits need updating, and proper tuning can significantly reduce latency and boost agent task completion rates.

Confirmed

• Thinking levels: On the medium level, Opus 5.5 matches or even exceeds Opus 5's high-level performance on many tasks; the guide recommends medium by default, low when speed matters, and xhigh/max only where measured gains justify it.
• Thinking can't be disabled: Thinking is always on and can no longer be turned off; old prompts with instructions like "think step by step" or "show your work" that force exposure of internal reasoning should be removed—start from low effort instead.
• Latency optimization: In Anthropic's testing, removing "think deeply before answering"-style instructions in a chat app cut first-token latency from about 8.2 seconds to about 2.1 seconds with no noticeable quality drop.
• Agents stopping mid-task: Claude may report progress and end its turn before finishing long tasks; the remedy is maintaining a task checklist, checking for unfinished items before declaring completion, auto-resuming when needed, and capping retries to avoid infinite loops.
• Progress visibility: Long-running agents can appear frozen; enable progress updates in your API integration and require brief check-ins at key milestones (e.g., 25%, 50%, 75%).
• Multi-app tasks: When working across Gmail, Docs, Sheets, and CRM, prompt Claude to check relevant connected data sources before acting—this measurably improves cross-app task completion rates.
• Time budgets: Telling a multi-agent system how much time has elapsed and the total estimated budget encourages parallel work; in one example, a 30-minute serial task shrank to roughly 15 minutes in parallel.
• Handling refusals: Opus 5.5 has guardrails around biosafety, cybersecurity, and requests to extract internal reasoning; when refused, first check the category rather than treating it as a technical failure, and use supported summarized-output options instead.
• Prompt injection defense: Wrap pasted content from emails, documents, webpages, etc. in clearly named tags (e.g., externalcontent) and explicitly instruct Claude to treat it as external content.
• Complex charts: Provide high-resolution images and give Claude tools to crop, zoom, and inspect details (PIL is officially recommended).
• Design constraints: Vague asks like "make a beautiful website" or "don't look AI-generated" work poorly; give specific constraints, such as avoiding cream-colored backgrounds, italic headline flourishes, numbered section labels, and pill-shaped buttons.

Why it matters

This guide directly affects Opus 5.5's real-world cost and experience: always-on thinking means toggle-style prompting strategies from older models no longer apply, and picking the wrong level or keeping redundant thinking instructions adds seconds of latency penalty; meanwhile, the checklist, progress, and time-budget techniques for agent scenarios offer an officially validated engineering playbook for building reliable multi-agent systems.

2026-10-28 ~ 2026-10-29 · 25 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
• 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
• 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
• [source] Anthropic releases Claude Opus 5.5 prompting guide with 11 best practices — FinanceYF5 · 2026-10-29
• Anthropic's Opus 5.5 prompting guide: don't max out thinking effort on every task — FinanceYF5 · 2026-10-29
• Opus 5.5 thinking is always on: medium effort rivals high-effort Opus 5, says Anthropic — FinanceYF5 · 2026-10-29
• Stop Claude agents quitting halfway: task checklist plus retry limit — FinanceYF5 · 2026-10-29
• Opus 5.5 refusals are safety guardrails, not bugs — check the refusal category — FinanceYF5 · 2026-10-29
• Anthropic guide: make long-running Claude agents show progress updates at milestones — FinanceYF5 · 2026-10-29
• Anthropic: prompt Claude to search connected sources before acting to boost task completion — FinanceYF5 · 2026-10-29
• Give multi-agent teams a time budget: Anthropic says agents work faster in parallel — FinanceYF5 · 2026-10-29
• Anthropic: removing 'think carefully' prompts cut Opus 5.5 latency from 8.2s to 2.1s — FinanceYF5 · 2026-10-29
• Opus 5.5 guide #9: wrap pasted content in tags to guard against prompt injection — FinanceYF5 · 2026-10-29
• Opus 5.5 guide #10: use high-res images and PIL/OpenCV tools for complex visuals — FinanceYF5 · 2026-10-29
• More Opus 5.5 prompting tips: hi-res images for charts, specific design constraints for sites — FinanceYF5 · 2026-10-29