Getting the Most Out of Opus 5.5

Opus 5.5 is built to fit into Claude the way people already use it, but it changes the rhythm. It works longer on its own, explains what it did in plain language, and always thinks before it replies. The practical advice from TLDR Dev is simple: stop micromanaging every turn and start giving it a clear finish line.

The biggest shift is how you ask. Give Opus 5.5 the full task in one message, spell out what “done” means, and tell it when to stop and ask for help. That matters most on multi-step work, where the model is said to do better than Opus 5 did, including long coding jobs that can stretch for hours with little oversight. The guide’s example is blunt: migrate the payment endpoints, delete the old client, and keep going until the tests pass.

Just as important, don’t waste words telling it to think harder. Opus 5.5 already thinks before every reply, so lines like “think carefully” and “think step by step” are mostly noise. The same goes for design prompts: instead of asking for a “non-generic” look, list the styles you don’t want, because the model falls back to its own defaults when left to guess.

For long runs in Claude Code, the article leans hard on structure. Put stop-and-go rules in CLAUDE.md, tell it when to keep going, keep a task list in a file, and use subagents for audits or migrations across a large codebase. When the run finishes, read what it needs from you first, because Opus 5.5 now calls out blockers and next steps more clearly than before. And if a message gets flagged, the system may switch to an older model; you can switch back, but that’s another reason to be explicit about what you want from the start.