Trade “think carefully” for a finish line
Trade staged, supervised prompts for a single, comprehensive task definition. Opus 5.5 now performs adaptive thinking, dynamically allocating its internal reasoning budget. Give it the entire task up front; then, define “done” with observable, objective criteria, letting the model execute autonomously.
Define “done” with explicit, verifiable checks. For a payment migration, this means:
- Every endpoint uses the new client
- The old client is deleted
- All tests pass
These specific finish lines eliminate ambiguity and allow Opus 5.5 to self-validate.
Remove generic reasoning cues like "think carefully" or "think step-by-step." Anthropic’s testing shows these phrases add latency without improving output quality. Opus 5.5 always thinks first and decides its own reasoning depth.
Instead, specify real constraints: "only stop if a test fails for a reason you cannot explain." This directs the model to resolve issues independently, pausing only when genuinely blocked or when an unexplainable test failure occurs. Vague instructions are ineffective; specific, actionable finish lines drive results.
Design gets better when you name what to avoid
Vague design prompts lead to generic output. Asking Opus 5.5 to “make it less generic” only swaps one template for another, not delivering distinctive results. The model defaults to familiar styles without specific direction.
Instead, define what to avoid. Better Stack’s video highlights effective negative constraints: no numbered sections, monospace labels, or pill-shaped buttons. AI models often favor these elements, so explicitly banning them forces more creative solutions.
Balance exclusions with positive guidance. Provide a clear target: describe the intended audience, the desired visual mood, hierarchy, and interaction needs. This gives Opus 5.5 a useful framework to generate a truly unique design.
For instance, specify: "Design a dashboard for senior executives, conveying immediate clarity and sophisticated simplicity. Prioritize data visualization over text. Avoid any playful or overly decorative elements. Ensure all interactive components are clearly distinguishable but visually understated." This combination of what not to do and what to achieve yields superior design outcomes from Opus 5.5.
Keep the run moving without losing control
Keep the run moving without losing control
Avoid costly restarts. If a new requirement surfaces mid-run, queue a follow-up directly in Claude Code. The harness processes the instruction at the next step, without interrupting the current operation or invalidating cached thinking. This preserves context and accelerates iteration.
Establish clear rules for agent autonomy. Add a CLAUDE.md rule: "keep going unless you need me or unless it's something destructive." This prevents Opus 5.5 from stalling on trivial updates, ensuring continuous progress while maintaining human oversight for critical actions.
For large-scale tasks like migrations or audits, delegate work to sub-agents. Instruct the main agent to split tasks, then verify each sub-agent's evidence before acceptance. This strategy reduces context for individual agents, improving accuracy and providing a multi-step review process.
Maintain a tasks.md file for long-running operations. Opus 5.5 summarizes older turns in the context window, potentially obscuring details. A persistent tasks.md preserves the complete task list, showing what's done and what remains, ensuring no progress is lost. For additional guidance, refer to the Prompting Best Practices - Claude Platform Docs.
Enjoying this? Get one like it in your inbox each morning.
one email a day · unsubscribe in two clicks · no third-party tracking
Review the diff—and watch for a model switch
Final report reviewed? First, check the “needs from you” section. Request a summary with “blocked on me,” “changed,” and “found” sections to clarify next steps. Opus 5.5 is good at self-auditing; leverage it.
Before accepting changes, request a merge-blocking-only diff review. Specify file, line, defect explanation, and reproduction steps. For independent verification, consider using a different model like Astra to review Opus’s output; varying models often yield more unique insights.
Opus 5.5 integrates fable-level bio and cyber safeguards, making it more cautious than Opus 5. These safeguards can flag legitimate security work on your own codebase. When flagged, ensure prompts clearly state you’re targeting your own code and avoid requests for internal reasoning, as this is a flag category.
Check your model-switch settings. Anthropic allows you to choose whether Claude automatically switches models or asks you first when a flag is triggered. If a prior message keeps triggering the safeguard, start a new thread to clear the context.
Frequently Asked Questions
Should you tell Opus 5.5 to think step by step?
Usually not. The guide says generic instructions such as “think carefully” add little; give the model clear requirements and success criteria instead.
How can you keep Claude Code from stopping mid-task?
Add a rule to CLAUDE.md describing when to continue and when to ask for help—for example, continue unless blocked or about to take a destructive action.
Can you add instructions while Claude Code is running?
Yes. Queue a follow-up while the run is active rather than interrupting and restarting the task.
Why might Claude switch to an older model?
A message may trigger a safety safeguard. Clarify legitimate context, review model-switch settings, and consider starting a fresh thread if the issue persists.