LLM Wiki

Appearance

Built 26/04/18 17:24commit 5ccb4ff

Best practices for using Claude Opus 4.7 with Claude Code ​

中文 | English

Summary ​

This Anthropic post is the practical companion to the Opus 4.7 launch note. Instead of announcing the model, it explains how operators should retune Claude Code around it: stronger first-turn briefs, fewer user interruptions, xhigh as the default effort, adaptive thinking instead of fixed-budget thinking, and more explicit instructions when you want aggressive tool use or parallel subagents.

Source ​

Key Contributions ​

  • Reframes interactive Claude Code use as delegation to a capable engineer, not tight turn-by-turn pair programming.
  • Makes first-turn task packaging operationally central: intent, constraints, acceptance criteria, and file locations should arrive early instead of being dripped in over many turns.
  • Explains why Opus 4.7 can consume more tokens in interactive sessions, especially because it reasons more after user turns.
  • Establishes xhigh as the recommended default effort for most coding and agentic tasks, with max reserved for intentionally expensive edge cases.
  • Records Anthropic's shift from fixed-budget Extended Thinking to adaptive thinking in Opus 4.7.
  • Notes a behavior shift toward fewer default tool calls and fewer default subagent spawns, which means harness prompts may need to be more explicit about when exploration or fan-out is desired.

Practical Implications ​

  • Prompt quality now depends even more on whether the first turn already contains the full task contract.
  • Interactive sessions should be batched more aggressively; fragmented operator guidance wastes tokens and can reduce overall quality.
  • Teams that previously relied on Opus 4.6 defaults may need to restate tool-use and delegation expectations in repo instructions, harness prompts, or skills.
  • Effort tuning is now a first-class Claude Code operating control, not a secondary optimization knob.
  • Completion notifications and auto mode belong in the same operational toolkit as prompt design because they reduce supervision overhead on long runs.

Relationship To Existing Anthropic Material ​

This page pairs naturally with Introducing Claude Opus 4.7:

  • the launch note explains what changed in the model and control surface
  • this post explains how to actually operate Claude Code differently because of those changes

Taken together, they show that model release notes and harness usage guidance now have to be read as one continuous operator surface.