Claude Opus 4.7 — The 10 Migration Moves
Claude Opus 4.7 shipped April 16, 2026. The behavioral shift from 4.6 is deliberate: 4.6 guessed intent when instructions were ambiguous; 4.7 takes instructions literally and will not silently generalize. Prompts that relied on the model "figuring out what you meant" break.
The Core Behavioral Shift
4.6 filled gaps when instructions were unclear. 4.7 takes you at your word. For products and agents that relied on implicit inference, this appears as a regression. For precise agentic work, this is the intended improvement.
Anthropic's migration guide: 4.7 "takes the instructions literally" and "will not silently generalize."
Boris Cherny (Claude Code lead, Anthropic): "It took a few days for me to learn how to work with it effectively."
The Governing Principle: Intent Engineering
4.7 rewards clear intent. Everything else in the migration is a tactic in service of this principle.
Intent has two layers:
Strategic context — durable. What you're building, who it's for, what's off-limits, what good looks like. Write this once in CLAUDE.md. It loads every session, so you stop retyping the project context on turn one.
Per-task intent — variable. What specifically to do right now. Written each turn. The gain from CLAUDE.md is removing the "remember what we're building" overhead from every task prompt.
See also claude-code/pawelhuryn-claude-decision-engine on using CLAUDE.md as a decision engine.
The 10 Migration Moves
1. Front-load intent in CLAUDE.md
Strategic context belongs in CLAUDE.md, not in every prompt. Delegate to secondary files (e.g., strategy.md) for content the agent doesn't need to read every session.
2. Default to Extra high (xhigh) effort
4.7 introduces a new effort level between high and max. Anthropic recommends xhigh for coding and agentic work. max is prone to overthinking — most "4.7 feels slow" complaints trace to running max by default. Use max only when the problem actually requires deep reasoning.
3. Toggle effort mid-task
Effort is per-call, not per-session. Use max for hard sub-problems; drop back to high or xhigh for the rest. No need to commit a single level to the whole session.
4. Regression-test old prompts
New tokenizer in 4.7: 1.0–1.35× more tokens per input. Existing 4.6 workflows cost more before any changes. Anthropic raised rate limits alongside the 4.7 launch. What matters: cost per correct output token, not tokens per prompt.
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.
6. Show what you want; remove prohibitions
Positive examples outperform negative rules. Per Anthropic guidance:
"Like this: [example]"works."Don't do this: [example]"burns tokens and rarely lands.
If a prompt has more than three "don't" or "never" lines, flip them: what does the ideal output look like? Two examples plus zero prohibitions is better.
7. Delete old progress scaffolding
Remove prompts like "summarize every 3 tool calls" or "give me a status update before moving on." 4.7 emits high-quality progress updates natively in long agentic traces. These scaffolds are now clutter that costs tokens.
8. Tell it to fan out explicitly
4.7 spawns fewer subagents by default and makes fewer tool calls per task. For parallel exploration, state it: "spawn subagents in the same turn to investigate X, Y, Z." Autonomy went up; default delegation went down.
9. Review plans, not diffs
Two primitives:
- Plan mode (Shift+Tab twice in Claude Code CLI) — surfaces the plan before code exists in the current session. Use for any change touching more than one file.
- /ultraplan (CLI only) — cloud-based plan drafting, review in browser while terminal stays free.
Because 4.7 takes intent literally, a small misread in the plan compounds into a large misread in the diff. Reviewing a 10-line plan for intent drift takes 30 seconds; reviewing a 200-line diff takes 15 minutes.
10. Switch to adaptive thinking
Fixed thinking budgets (budget_tokens) are removed. API calls with budget_tokens return HTTP 400. Replacement:
thinking={"type": "adaptive"}
Use alongside the effort parameter. Simple completions get minimal thinking; hard problems scale up automatically.
Benchmark Profile
4.7 shows improvements in coding, creative writing, and structured work. Instruction following on vague prompts, multi-turn conversations, and long-context retrieval shows regressions versus 4.6 on some benchmarks. These are traded capabilities, not uniform upgrades. The trade is intentional — precision over inference.
Convergence Pattern
Anthropic moved 4.7 toward more literal instruction following. OpenAI's December 2025 Model Spec updated to "consider not just the literal wording but the underlying intent." The two leading labs are converging from opposite directions. Intent engineering is the transferable skill that works on both.
Related
- claude-code/session-management-guide — session management and context optimization
- claude-code/token-optimization-toolkit — token cost management
- claude-code/pawelhuryn-claude-decision-engine — CLAUDE.md as decision engine
- claude-code/ultraplan — ultraplan feature
- claude-code/context-window-optimization — context optimization