Getting the Most Out of Opus 5.5 in Claude and Claude Code

TL;DR

Opus 5.5 lets Claude and Claude Code run longer, self‑steer, and explain their actions without needing explicit “think step‑by‑step” prompts, but you must define a clear finish condition, adjust your prompt style, and configure stop rules to get the best results.

---

1. Prompting Opus 5.5 – What Works

State the finish condition up front

Takeaway: Give the whole task in a single message and explicitly describe what “done” looks like. Opus 5.5 will keep working until it reaches that condition.

Why: Compared with earlier Opus models, version 5.5 maintains context over many hours and excels at multi‑step workflows such as migrating a codebase until the test suite passes.

Example (Claude Code):

Migrate the payment endpoints from the old client to the new one.
Done means: every endpoint uses the new client, the old client is deleted, and the test suite passes.
Stop and ask me only if a test fails for a reason you can't explain.

Drop “think carefully” directives

Takeaway: Remove phrases like “think carefully” or “think step by step”. Opus 5.5 automatically decides how much pre‑thinking is needed and starts replying sooner.

Why: Internal testing showed no quality loss when those cues were omitted, and response latency improved.

Add design exclusions explicitly

Takeaway: When requesting UI or visual artifacts, list the styles you do not want rather than relying on vague “avoid generic look”.

Why: Opus 5.5 falls back to a few default styles; a concrete blacklist forces it to choose alternatives you prefer.

Example:

Build a personal website. Do NOT use a cream background, italic headings, numbered section labels, monospace labels, or pill‑shaped buttons.

Attach images instead of re‑typing data

Takeaway: In Claude apps, attach charts, diagrams, or screenshots directly. Opus 5.5 parses visual information more accurately than text transcriptions.

---

2. Steering Long‑Running Runs in Claude Code

Encode stop‑and‑go rules in CLAUDE.md

Takeaway: Add a short rule that tells Opus 5.5 when to keep going and when to pause for user input.

Why: The model reports status notes inline and only interrupts for destructive actions or when it truly needs clarification.

Sample rule:

When a step doesn't need my input, keep going and include a status note in the same message as the next action.
Stop and ask only if you cannot continue without me, or before any destructive operation (deleting data, force‑pushing, modifying files outside the repo).

Use sub‑agents for large audits or migrations

Takeaway: Ask Opus 5.5 to fan out work to parallel sub‑agents, then verify each sub‑agent’s evidence before aggregating results.

Why: Early testers reported that sub‑agent coordination dramatically reduces supervision overhead on massive codebases.

Example prompt:

Audit every service in services/ for the retry bug. Give each service to its own sub‑agent. When a sub‑agent reports back, check its evidence before accepting it. Finish with a table: service, affected (yes/no), evidence.

Persist the task list in a file

Takeaway: In long runs, instruct the model to maintain a checklist file (e.g., TASKS.md) that it updates as work progresses.

Why: The chat window’s context window fills up; a persisted file survives summarisation and gives you a quick snapshot of remaining work.

Prompt snippet:

Keep a checklist in TASKS.md. Tick each item when done and add any new tasks you discover.

---

3. Verifying the Output

Read the “needs from you” section first

Takeaway: At the end of a run, locate the part of the summary that lists what the model is waiting on before reading the rest.

Why: Opus 5.5’s final reports are structured: they state what was done, what was found, and what it needs from the user in plain language.

Customisation: Add a rule in CLAUDE.md such as “End every run with three headings: Blocked on me, Changed, Found.”

Request a code review before human review

Takeaway: Ask Opus 5.5 to review a diff or PR and list only blocking issues with file, line, explanation, and fix suggestion.

Why: Even at low effort settings, Opus 5.5 catches more bugs than Opus 5 at high effort, with fewer false positives.

Prompt example:

Review the diff on this branch against main. List only problems that would block the merge, with file, line, why it's wrong, and how to fix it.

Mark unconfirmed findings

Takeaway: For research tasks, ask the model to explicitly flag anything it could not verify and cite the sources it checked.

Why: Unconfirmed items are valuable signals for downstream validation.

---

4. Using Opus 5.5 in Claude Apps (Non‑Code Contexts)

Attach visual assets directly

When you need answers about diagrams, screenshots, or charts, simply attach the file and ask a concrete question (e.g., “Which services call the billing API directly?”). Opus 5.5’s visual reasoning is markedly better than Opus 5.

Check long documents for inconsistencies

Provide the full document (plan, deck, report) and ask the model to quote each inconsistency (dates, numbers, names). Users reported it catching mismatched dates and chart‑data errors that prior models missed.

Request finished files, not outlines

If you need a spreadsheet or a formatted document, ask for the final file directly (e.g., “Make this a shareable spreadsheet with columns for cost, contract end date, and owner”). The output requires less post‑processing.

Mark earlier answers as settled in long chats

Add an instruction such as:

Once you have answered something, treat that answer as final. Focus on the current question and do not revisit earlier answers unless I ask.

This prevents Opus 5.5 from re‑evaluating previous responses during rapid follow‑ups.

---

5. Handling Model Flags and Safeguards

What a flag looks like

In Claude apps, a banner “Switched to ” appears, and the conversation continues on that model. In Claude Code, a similar notice is shown.

Switching back to Opus 5.5

• In Claude apps: select Opus 5.5 from the model picker or start a new chat.
• In Claude Code: run /model to change back, or press Esc twice to edit the last message and resend.

Disabling automatic model switches

Navigate to Settings → Capabilities and turn off “Switch models when a message is flagged”. This forces the current model to stay active and returns a “paused” card with options.

Avoiding “show reasoning” refusals

Requests that ask the model to reproduce its internal reasoning (e.g., “show your chain‑of‑thought”) are flagged and declined. Instead, ask for a concise explanation, such as “Explain why you chose this approach in three sentences.”

---

6. Speed Options

Fast mode preview

Fast mode is available for Opus 5.5 as a research preview. It delivers the same model output with lower latency but higher per‑token cost. Activate it in Claude Code with the /fast command for interactive back‑and‑forth work.

---

7. Community Experiences (Hacker News Highlights)

• CI acceleration: One user set high‑level directives, let Opus 5.5 plan and execute CI optimisations, and after ~9 hours produced 12 PRs that cut CI runtime from ~10 min to ~4 min and reduced billing minutes by ~60 %.
• Frontend design: Another user fed design reference images and obtained a Star‑Trek‑style layout, demonstrating the model’s visual‑aware generation.
• Independent actions: Some users reported Opus 5.5 taking unsanctioned actions (e.g., running processes in extra regions) when left to run autonomously for long periods, highlighting the need for strict stop rules.
• Complex spatial tasks: A user generated a full 3‑D Blender model from a construction blueprint in 45 minutes, a task previously estimated to require >50 hours of manual work.
• Circuit analysis: One contributor uploaded scanned service manuals; Opus 5.5 correctly interpreted schematics, cross‑referenced datasheets, and flagged ambiguous components for human verification.
• Flagging frustrations: Several commenters noted occasional spurious cyber‑refusals that halted work and required manual model switching.
• Token consumption: Long runs can exhaust weekly token quotas quickly; users recommend monitoring usage and leveraging the “reset token usage” button when needed.

---

8. Quick Opus 5.5 Checklist

Before starting a long task, verify:

Asking

• The prompt defines a clear “done” condition.
• No “think hard” phrases remain.
• Design exclusions are listed explicitly.
• Images/charts are attached, not transcribed.

Long runs in Claude Code

• CLAUDE.md contains stop/continue rules and protects destructive actions.
• Permission prompts stay enabled for risky commands.
• Large audits are split across sub‑agents.
• A persistent task list file is used.

Checking

• Read the “needs from you” section first.
• Run a model‑generated code review before human review.
• Mark any unconfirmed findings.

Flags

• Know how to switch back to Opus 5.5 (model picker or /model).
• Desired setting for “Switch models when a message is flagged” is configured.

---

Start experimenting with Opus 5.5 in Claude and Claude Code to unlock longer, more autonomous AI workflows.