中文 | English

Best practices for using Claude Opus 4.7 with Claude Code

Source: https://claude.com/blog/best-practices-for-using-claude-opus-4-7-with-claude-code Fetched via web on 2026-04-19 for llm-wiki ingest.

Summary

This Anthropic post explains how to retune Claude Code usage for Opus 4.7. The core message is that Opus 4.7 should be treated less like a turn-by-turn pair programmer and more like a capable engineer who benefits from a strong first-turn brief, fewer interruptions, explicit effort tuning, and clearer instructions about when to use tools or spawn subagents.

Structuring interactive coding sessions

Opus 4.7 behaves differently in interactive multi-turn sessions versus more autonomous asynchronous runs. In interactive settings, it reasons more after user turns, which improves coherence and coding quality over long sessions but also increases token usage.

Anthropic recommends:

• specify the task clearly in the first turn, including intent, constraints, acceptance criteria, and relevant file locations
• reduce the number of required user interactions, because every user turn adds reasoning overhead
• use auto mode when the model can proceed safely without frequent check-ins
• set up completion notifications so the operator can step away while the task runs

Recommended effort settings for Opus 4.7

The default effort level in Claude Code for Opus 4.7 is now xhigh, positioned between high and max.

Anthropic's guidance:

• low and medium for cost-sensitive, latency-sensitive, or tightly scoped work
• high when balancing intelligence and cost, especially across concurrent sessions
• xhigh as the default and recommended setting for most coding and agentic work
• max only for genuinely hard and non-cost-sensitive tasks, because it has diminishing returns and is more prone to overthinking

The post explicitly recommends experimenting with effort rather than carrying over an older setting unchanged from Opus 4.6.

Working with adaptive thinking

Extended Thinking with a fixed thinking budget is not supported in Opus 4.7. Instead, the model uses adaptive thinking, deciding when to spend more reasoning based on context.

Anthropic suggests prompting directly if you want more or less thinking, for example:

• ask it to think carefully and step by step when the problem is harder than it looks
• ask it to prioritize speed over deep thinking when you want lower token usage and can accept some accuracy tradeoff

Behavior changes worth knowing

Compared with Opus 4.6, Anthropic says Opus 4.7:

• calibrates response length more to task complexity, instead of being default-verbose
• calls tools less often and reasons more unless tool use is explicitly encouraged
• spawns fewer subagents by default and should be told more directly when parallel delegation is desirable

The post gives a concrete subagent instruction pattern: do not spawn a subagent for work that can be completed directly in one response, but do spawn multiple subagents in the same turn when fanning out across items or files.

Why this matters

This is not just a prompting note. It is an operator-facing harness retuning guide. Anthropic is effectively saying that model upgrades change optimal first-turn briefing style, turn structure, effort defaults, tool-use prompting, and delegation policy inside Claude Code.