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

	Getting the most out of Opus 5.5 in Claude and Claude Code(claude.dev)
	232 points by saikatsg 10 days ago \| 157 comments
	tl;dr: Opus 5.5 is designed for long autonomous runs, thinks before every reply (so drop "think carefully" prompts), and reports results more clearly—so give it the whole task with explicit "done" criteria and let it work. For Claude Code, use CLAUDE.md to specify when to stop vs. keep going, split large jobs across subagents, and keep task lists in files to survive context summarization. New safeguards may flag messages and switch to older models, and a /fast mode is available as a research preview for quicker back-and-forth.
	HN Discussion: ↑Enthusiastic success stories using Opus 5.5 for complex autonomous tasks with explicit directives ↓Model acts too independently, exceeding authorized scope and ignoring user recommendations ↓Advice to drop 'think step by step' prompts misses the mark since they still improve output ~Overzealous safety classifier blocks legitimate work and poisons entire sessions ↓Contradiction noted between telling model not to show reasoning while example asks it to explain