claude-opus-4-5-migration

作者: anthropic

將 Claude 模型字串與提示從 Sonnet 4.0、Sonnet 4.5 或 Opus 4.1 遷移至 Opus 4.5。更新 Anthropic API、AWS Bedrock、Google Vertex AI 及 Azure AI Foundry 平台上的模型字串。移除不支援的測試版標頭(例如 context-1m-2025-08-07),並新增設為「高」的 effort 參數。針對已知的 Opus 4.5 行為差異提供可選的提示調整:工具過度觸發、過度工程化、程式碼探索保守性、前端設計及思考敏感度...

npx skills add https://github.com/anthropics/claude-code --skill claude-opus-4-5-migration

Opus 4.5 Migration Guide

One-shot migration from Sonnet 4.0, Sonnet 4.5, or Opus 4.1 to Opus 4.5.

Migration Workflow

  1. Search codebase for model strings and API calls
  2. Update model strings to Opus 4.5 (see platform-specific strings below)
  3. Remove unsupported beta headers
  4. Add effort parameter set to "high" (see references/effort.md)
  5. Summarize all changes made
  6. Tell the user: "If you encounter any issues with Opus 4.5, let me know and I can help adjust your prompts."

Model String Updates

Identify which platform the codebase uses, then replace model strings accordingly.

Unsupported Beta Headers

Remove the context-1m-2025-08-07 beta header if present—it is not yet supported with Opus 4.5. Leave a comment noting this:

# Note: 1M context beta (context-1m-2025-08-07) not yet supported with Opus 4.5

Target Model Strings (Opus 4.5)

PlatformOpus 4.5 Model String
Anthropic API (1P)claude-opus-4-5-20251101
AWS Bedrockanthropic.claude-opus-4-5-20251101-v1:0
Google Vertex AIclaude-opus-4-5@20251101
Azure AI Foundryclaude-opus-4-5-20251101

Source Model Strings to Replace

Source ModelAnthropic API (1P)AWS BedrockGoogle Vertex AI
Sonnet 4.0claude-sonnet-4-20250514anthropic.claude-sonnet-4-20250514-v1:0claude-sonnet-4@20250514
Sonnet 4.5claude-sonnet-4-5-20250929anthropic.claude-sonnet-4-5-20250929-v1:0claude-sonnet-4-5@20250929
Opus 4.1claude-opus-4-1-20250422anthropic.claude-opus-4-1-20250422-v1:0claude-opus-4-1@20250422

Do NOT migrate: Any Haiku models (e.g., claude-haiku-4-5-20251001).

Prompt Adjustments

Opus 4.5 has known behavioral differences from previous models. Only apply these fixes if the user explicitly requests them or reports a specific issue. By default, just update model strings.

Integration guidelines: When adding snippets, don't just append them to prompts. Integrate them thoughtfully:

  • Use XML tags (e.g., <code_guidelines>, <tool_usage>) to organize additions
  • Match the style and structure of the existing prompt
  • Place snippets in logical locations (e.g., coding guidelines near other coding instructions)
  • If the prompt already uses XML tags, add new content within appropriate existing tags or create consistent new ones

1. Tool Overtriggering

Opus 4.5 is more responsive to system prompts. Aggressive language that prevented undertriggering on previous models may now cause overtriggering.

Apply if: User reports tools being called too frequently or unnecessarily.

Find and soften:

  • CRITICAL: → remove or soften
  • You MUST...You should...
  • ALWAYS do XDo X
  • NEVER skip...Don't skip...
  • REQUIRED → remove or soften

Only apply to tool-triggering instructions. Leave other uses of emphasis alone.

2. Over-Engineering Prevention

Opus 4.5 tends to create extra files, add unnecessary abstractions, or build unrequested flexibility.

Apply if: User reports unwanted files, excessive abstraction, or unrequested features. Add the snippet from references/prompt-snippets.md.

3. Code Exploration

Opus 4.5 can be overly conservative about exploring code, proposing solutions without reading files.

Apply if: User reports the model proposing fixes without inspecting relevant code. Add the snippet from references/prompt-snippets.md.

4. Frontend Design

Apply if: User requests improved frontend design quality or reports generic-looking outputs.

Add the frontend aesthetics snippet from references/prompt-snippets.md.

5. Thinking Sensitivity

When extended thinking is not enabled (the default), Opus 4.5 is particularly sensitive to the word "think" and its variants. Extended thinking is enabled only if the API request contains a thinking parameter.

Apply if: User reports issues related to "thinking" while extended thinking is not enabled (no thinking parameter in request).

Replace "think" with alternatives like "consider," "believe," or "evaluate."

Reference

See references/prompt-snippets.md for the full text of each snippet to add.

See references/effort.md for configuring the effort parameter (only if user requests it).

來自 anthropic 的更多技能

access
anthropic
管理 Discord 頻道存取權限 — 核准配對、編輯允許清單、設定私訊/群組政策。當使用者要求配對、核准某人、查詢誰被允許時使用…
official
session-report
anthropic
從 ~/.claude/projects 的對話記錄中,生成一份可探索的 HTML 報告,內容涵蓋 Claude Code 工作階段的使用情況(包含 token、快取、子代理、技能及高成本提示)。
official
build-mcp-server
anthropic
當使用者要求「建立 MCP 伺服器」、「建立 MCP」、「製作 MCP 整合」、「為 Claude 包裝 API」、「暴露工具給…」時,應使用此技能。
official
cookbook-audit
anthropic
根據評分標準審核 Anthropic Cookbook 筆記本。每當要求進行筆記本審查或審核時使用。
official
handle-complaint
anthropic
處理進線客戶投訴的完整流程——提取背景資訊、草擬回覆,並建議營運改善方案。可接受選填的電子郵件或工單編號……
official
use-case-triage
anthropic
快速判斷某項處理活動是否需要PIA、強制性的GDPR DPIA,或可直接進行——揭露隱私政策衝突,並引導至正確的…
official
board-minutes
anthropic
根據您指定的格式,自動草擬董事會或委員會會議記錄。自動從您的日曆中偵測即將召開的董事會及委員會會議,詢問議程及…
official
renewal-tracker
anthropic
顯示即將到來的取消截止日期的合約,並在通知窗口關閉前發出警告,依據維護中的續約登記表進行操作。當用戶詢問時使用…
official