Anthropic ships official Opus 5.5 prompting guide: 12 practical tips distilled

BuffaloConscious7919 · reddit · 2026-09-28

Anthropic released an official prompting guide for Claude Opus 5.5; a Redditor distilled it into 12 key points:

Time budgets for agent teams: markers like "elapsed 340s / 1200s" help agent teams pace themselves and finish sooner
Progress updates now arrive as thinking blocks (empty by default) between tool calls — text-only UIs may look frozen
Thinking can't be disabled: migrating from Opus 5 with thinking off? Start at low effort and measure
Unattended agents stop early: some progress updates end the turn; loops treating that as "done" quit mid-task
New guardrails: safety classifiers cover biology, cybersecurity and reasoning extraction; avoid asking it to write out its reasoning
Tag pasted text with random IDs so hidden instructions in copied content get ignored
Look around first: asking it to check emails/docs/records across connected apps before acting improves results
Better image reading for charts, diagrams and screenshots
Drop "think carefully" — faster replies, no clear quality drop
Fix generic AI look by naming exact styles to ban instead of "avoid generic"
Effort is the main dial: medium default matches or beats Opus 5 on high; lowering effort beats prompt instructions
Old prompts still work: 30%+ faster than Opus 5, fewer tokens

Also: changing top-level effort invalidates the prompt cache; use per-message effort (beta) to keep it.

More from coding & agent