Anthropic has released a detailed usage guide for Claude Opus 5.5, and the document reads less like a changelog and more like a field manual for a fundamentally different working relationship between human and AI. The core shift: Opus 5.5 runs longer on its own, decides how much to think before every reply, reports plainly what it did, and can coordinate parallel subagents across large codebases. Users are explicitly told to hand over whole tasks, define what "done" looks like, and then step back. The most revealing detail is the negative advice. Anthropic tells users to delete "think carefully" and "think step by step" prompts — staples of the prompt-engineering era — because Opus 5.5 already thinks before every reply and calibrates depth autonomously. In testing, removing these lines made replies start sooner with no quality drop. This is a quiet burial of an entire prompting paradigm that shaped how millions interact with language models. For Claude Code users, the operational model has shifted from pair programming to project management. The guide recommends writing a CLAUDE.md file that defines when the model should stop and ask versus keep going — essentially a set of standing orders. Users can add instructions mid-run without restarting, ask the model to fan work out across subagents for large audits or migrations, and have it maintain its own task list in a persistent file that survives context-window summarization. Early testers reportedly ran long coding tasks for hours with little oversight. The design guidance is unexpectedly specific. Anthropic acknowledges that Opus 5.5 defaults to a handful of visual styles when building pages or artifacts — cream backgrounds, italic accent words, pill-shaped buttons, numbered section labels. Rather than asking for "something unique," users are told to enumerate the patterns they want excluded, then iterate on whatever the model chooses instead. It is a frank admission that the model has aesthetic defaults, and the workaround is constraint-based rather than inspiration-based. On quality assurance, the guide positions Opus 5.5 as a pre-human reviewer. One early tester reportedly found that Opus 5.5 at its lowest effort setting caught more bugs than Opus 5 at high effort, with fewer false positives. The model is also better at reading charts, diagrams, and screenshots — understanding spatial relationships like which boxes an arrow connects or when a meeting starts in a calendar. For research tasks, users can ask it to explicitly mark what it couldn't confirm and where it looked, surfacing epistemic gaps rather than papering over them. The safety section introduces a notable trade-off. Opus 5.5 is the first Opus model with "Fable-level" bio and cyber safeguards. When a message is flagged, most work silently moves to an older model. Finding security vulnerabilities in source code is allowed, and everyday health questions should work, but Anthropic acknowledges these safeguards sometimes flag legitimate work and says they are tuning to reduce false positives. The practical implication: users doing security research or adjacent work may hit invisible model switches mid-session. Read together, these instructions describe a product that has crossed a threshold from tool to autonomous worker. The human's job is no longer to guide each step but to define outcomes, set boundaries, and review results. That is a genuine workflow shift — and the friction will come not from the model's capabilities but from users and organizations learning to trust a delegation pattern they have never practiced with software before.
