Prompting Claude Opus 5.5

Who Should Read

Agent developer moving from Opus 5 to Opus 5.5 who must rework effort settings, unattended agent loops, and progress update handling

Core Mechanics

• Medium effort is now the default
  • Opus 5.5 defaults to medium effort, while Opus 5 defaulted to high.
  • In Anthropic's testing, 5.5 at medium matched or beat Opus 5 at high on coding and knowledge-work evals.
  • At the same level 5.5 thinks more per turn, so carrying over old settings can mean longer turns and more tokens.
• Thinking is always on
  • 5.5 does not accept thinking: disabled.
  • Thinking tokens count toward max_tokens, so leave enough room.
  • For long agentic coding turns, the 128,000 maximum worked well.
• Unattended agents can stop midway
  • On long tasks the model may write a progress report and end the turn with end_turn.
  • A loop that treats this as completion stops there.
  • Use a checklist to detect open items and cap automatic continuations at two or three.
• Asking for reasoning in output can be refused
  • Requests that push the model to reproduce reasoning in the response may be declined with the reasoning_extraction category.

Community Reactions

• Evals are costly, one commenter said
  • Testing several effort levels on your own evals is right, but statistically significant benchmarking is expensive given LLM variance.
• Pushback on the progress-update harness
  • Commenters said most harnesses surface tool calls as progress, so good UI should be enough.
  • One suggested feeding a per-turn transcript to a cheaper model at fixed intervals instead.
• Interest in marking pasted text
  • One commenter found it interesting that pasted text is wrapped in tags with random IDs to resist prompt injection.
  • They had been skeptical of such protections but said a trained model might make it work.
• Real-world reports were mixed
  • One user said the output is verbose and the Concise style setting had no effect.
  • One said it roamed through directories outside the project on first run and never explained why.
  • Another ran a 33-hour session nearly unattended with good results.

How to Apply

• Set max_tokens and effort explicitly
  • Today, set effort to medium in code carried over from Opus 5 and raise max_tokens to cover thinking.
  • Run low, medium, and high on the same eval inputs and compare cost and quality.
• Change end_turn handling in unattended loops
  • Don't treat end_turn as done; if the checklist (todo tool or file) has open items, send a user message to continue.
  • Cap automatic continuations on the same task at two or three.
• Remove reasoning-output prompts and read by block type
  • Delete instructions like "write your reasoning in the answer" from prompts.
  • Set display: "summarized" and process content by block type instead of assuming the first block is text.
• Add an explore sentence for multi-app agents
  • Append the up-front exploration instruction to the system prompt of agents working across email, docs, sheets, and CRM.
  • Keep untrusted content out of the records, since the model acts on what it finds.

Try It

# Paste at the end of a Claude or ChatGPT conversation, or an agent's system prompt (for multi-app agents, based on the source sentence)

Before taking any action, explore broadly with tool calls: list and open the emails, documents, spreadsheet tabs and records across the available apps that could be relevant to this task, including ones the task does not explicitly mention, and use what you find.

# User message with the task: {{task description}}

Next step: run the same task with and without this sentence and compare success rate and tool call count

Terminology

Related Papers

• AI is a tool and using it well takes skill

  A rebuttal to Hickey: LLMs are tools, and skill in directing and checking them decides the outcome

• Reflective Prompt Optimization of a Deep-Research Agent: A Topic-Disjoint, Multi-Seed Protocol for Measuring Transferable Gains and Prompt-Level Overfitting

  Reflective prompt optimization raised held-out citation reward from 0.702 to 0.781 on a deep-research agent.

• GPT‑6 and Intelligent UI for everyone

  GPT-6 reportedly adds generated UI to chat; with no article text, only comment reactions are available.

• Base Models Can Reason By Taking a Cue From Training Data

  Prefilling just two opening tokens lifted Olmo-3-7B's MATH-500 accuracy from 42% to 78%.

• Labels Override Definitions in Jev-Style Typed Decision Models

  Renaming options to A/B raised routing accuracy by +0.1511 over meaningful labels.

• How Much Prompt Is Enough? A Blackbox Minimization of Few-Shots in LLMs

  Delta debugging trimmed few-shot prompts by 65.3% on average with unchanged propositional output

Related Resources

• https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-opus-5-5