🧭 "Building on Opus 4.6" — Developer Webinar Highlights

Anthropic held its Building on Opus 4.6 developer webinar today, drawing several thousand live attendees from the API developer community. The session was led by members of the developer experience and research teams and focused on practical guidance for production agentic applications built on Claude Opus 4.6. A recording is being made available to all registered participants.

Key guidance from the session

• System prompt structure matters at Opus scale — the team demonstrated that Opus 4.6 responds significantly better to hierarchical system prompts (role definition → task context → constraints → output format) than to flat prose instructions
• Extended thinking for decomposition — turning on extended thinking for the planning phase of a multi-step task, then switching to standard mode for execution, substantially reduces hallucinated tool calls
• Tool call batching — designing tool schemas to accept batched inputs (e.g., an array of file paths rather than a single path) yields 30–40% fewer round-trips in file-heavy workflows
• Cost guardrails — Anthropic recommends setting a max_tokens budget per sub-task rather than per top-level request when using Agent Teams, to prevent runaway generation in edge cases

🧭 Claude Code v2.1.50 — Token Counting Accuracy & /compact Command

Claude Code v2.1.50 has shipped with two improvements that address longstanding friction in long-running sessions. The first is a fix to token counting accuracy: the live token counter in the session header now correctly accounts for all tool-result payloads and cached context blocks, which previously caused the displayed count to diverge significantly from billing actuals on sessions with heavy file-read or web-fetch usage.

The second addition is a new /compact command that triggers an in-session context compaction: Claude summarises the conversation history to date into a concise state summary, discards the raw transcript, and continues the session with the summary as the new base context. This allows very long sessions to continue past the natural context limit without losing goal continuity. The compacted summary is saved to .claude/session-compact.md in the project root for review.