Claude Opus 4.7完全ガイド：4.6プロンプトを機能させる10の手順

The Ultimate Guide to Claude Opus 4.7 The ten moves to make your 4.6 prompts work on 4.7, leading with context, and why you don't need more instructions. 4.6 guessed. 4.7 stopped guessing. Your old prompts still work, mostly. The ones that break need one of the ten moves below. Anthropic shipped Claude Opus 4.7 on April 16. The official migration guide puts it plainly: 4.7 "takes the instructions literally" and "will not silently generalize." It's not a uniform upgrade. Wins on coding, creative writing, and structured work. Losses on instruction following in vague prompts, multi-turn, and long-context retrieval. The benchmarks show a trade, not a regression. Boris Cherny, Claude Code lead at Anthropic, posted on release day: "It took a few days for me to learn how to work with it effectively." ## 1. What's New in Opus 4.7 4.6 filled the gaps when your instruction was unclear. 4.7 takes you at your word. For PMs: if your 4.6 workflow relied on the model "figuring out what you meant," expect 4.7 to ask more questions, or to do less, or to do exactly what you asked for (which is not what you wanted). ## 2. Intent: The Universal Unlock This is the principle: 4.7 rewards clear intent. Everything else in this post is a tactic. Not longer prompts, not more rules, not a bigger CLAUDE.md. Intent splits into two layers: Strategic context is durable: what you're building, who it's for, what's off-limits, what good looks like. Write it once. Put it in CLAUDE.md. It loads every session, progressive-disclosure style, so you're not paying to reintroduce the project on turn one. Per-task intent is variable: what specifically do I want Claude to do right now. You still write this every turn. The gain from CLAUDE.md is that you stop retyping the strategic context on top of it. Treat Claude as a partner. Help it understand what and why you're doing, your strategy, objectives, and success criteria. Strategic context - not micromanaging the agent. ## 2.1 Anthropic and OpenAI convergence Anthropic moved Opus 4.7 toward more literal instruction following. OpenAI updated their December 2025 Model Spec to say "consider not just the literal wording but the underlying intent." They're converging from opposite directions. Anthropic is adding precision to its intent-first model. OpenAI is adding intent inference to its precision-first model. The same skill (engineering intent clearly) is becoming the unlock on both sides. ## 3. The 10 Claude Opus 4.7 Migration Moves ### 3.1 Front-load intent in CLAUDE.md You don't have to retype the strategic context every session. Put it in CLAUDE.md once. Every future session starts with the context already loaded. ### 3.2 Default to Extra high (xhigh) New effort level between high and max. Anthropic's own recommendation for coding and agentic work. max is prone to overthinking. Most "4.7 feels slow" reports trace back to people running max by reflex. Use max only when the problem actually needs deep reasoning. ### 3.3 Toggle effort mid-task Effort is per-call, not per-session. max for the hard subproblem. Drop back to high for the rest. ### 3.4 Regression-test old prompts New tokenizer. 1.0 to 1.35x more tokens per input. Your 4.6 workflows cost more on 4.7 before you've changed a line. ### 3.5 Batch questions. Stop drip-feeding. If you have three questions, ask all three in one message. On 4.6, clarifying across 3-4 turns worked. On 4.7, each turn adds reasoning overhead on top of literal interpretations from earlier turns. ### 3.6 Show what you want Positive examples beat negative rules. According to Anthropic: "Like this: " followed by short examples works. "Don't do this: " rarely lands and burns tokens trying. ### 3.7 Delete old progress scaffolding "Summarize every 3 tool calls." "Give me a status update before moving on." Delete these. Opus 4.7 emits high-quality progress updates natively in long agentic traces. ### 3.8 Tell it to fan out explicitly Opus 4.7 spawns fewer subagents by default and makes fewer tool calls per task. For parallel exploration, you now have to ask. Phrasings that work: "spawn subagents in the same turn to investigate X, Y, Z." ### 3.9 Review plans, not diffs Plan mode (Shift+Tab twice in the Claude Code CLI): inline, surfaces the plan before any code exists in the current session. /ultraplan (CLI only): cloud-based plan drafting from the CLI, review in the browser. Why this matters more on 4.7: because 4.7 takes intent literally, a small misread in the plan becomes a large misread in the diff. Reviewing a 10-line plan for intent drift takes 30 seconds. Reviewing a 200-line diff for the same drift takes 15 minutes. ### 3.10 Adaptive thinking only Fixed thinking budgets are gone. Use thinking: {type: 'adaptive'} plus the effort parameter. Old API calls with budget_tokens return HTTP 400. ```javascript client = anthropic.Anthropic() response = client.messages.create( model="claude-opus-4-7", max_tokens=16000, thinking={"type": "adaptive"}, messages=[ { "role": "user", "content": "Explain why the sum of two even numbers is always even.", } ], ) ``` ## 4. Closing 4.6 felt smart. 4.7 is honest. What looks like less intelligence is 4.7 refusing to guess. Your leverage is clear intent. Not reviewing every line of code, writing endless instructions, or retyping the same context every session. The more of your intent and expectations you push up to CLAUDE.md, the less you pay (in money and attention) to run the thing.