Playbooks for using Opus 5.5 in Claude and Claude Code

Ask for the whole task

A guide by Addy Osmani, published Sep 22, 2026, recommends giving Opus 5.5 the entire task in one message and defining completion, such as every endpoint being migrated or the tests passing. It describes the model as continuing longer on multi-step work than Opus 5, including large repository changes. Claude Code users can send a follow-up while a run is active instead of restarting it.

Opus 5.5 is described as thinking before every reply, so prompts do not need lines such as 'think carefully.' Simple questions can request a direct answer, and Claude Code's effort setting can change how much the model thinks. Design requests should name unwanted styles, including cream backgrounds, italic heading accents, numbered labels, monospace labels or pill-shaped buttons.

Steer long Claude Code runs

CLAUDE.md should state when a long run should keep going and when it should stop for input. The guide's example puts status notes with the next action and asks for help only when work is blocked or before destructive actions such as deleting data, force-pushing or changing files outside the repository. Permission prompts should remain enabled for destructive commands.

Large audits, migrations and reviews can be divided among subagents, with Opus 5.5 checking each result and its evidence. A checklist in TASKS.md can be updated during the run, giving users a current view of completed and remaining work even after older conversation turns are summarized.

Check results and inputs

After a run, users should read what Claude needs from them, including unresolved decisions or approvals, before the rest of the summary. The guide recommends asking Opus 5.5 to review a diff or pull request and list merge-blocking problems with a file, line, reason and failure demonstration. Research requests should mark anything the model could not confirm and say where it looked.

In Claude apps, charts, diagrams, screenshots and slides should be attached rather than retyped. Long plans, reports or decks can be checked for contradictions involving numbers, dates and names. Users should request finished spreadsheets or documents with the required structure. In long projects, earlier answers can be treated as settled unless a later step reveals a problem.

Handle model switches and speed

The guide says Opus 5.5 is the first Opus model to launch with Fable-level bio and cyber safeguards. Most flagged messages in Claude apps and Claude Code move to an older model while work continues. Users can switch back through the Claude apps model picker or with /model in Claude Code. Automatic switching can be changed in Settings or through /config, while /feedback can report an incorrect flag.

Claude Code's fast mode is available for Opus 5.5 at launch as a research preview. It uses the same model and delivers text sooner, but requires extra usage and costs more per token than standard mode. Users can enable it with /fast.