Paweł Huryn has published a detailed thread on Claude Opus 4.7, laying out what changed from Opus 4.6—and a set of practical “migration moves” aimed at teams whose older prompts now behave differently.
Anthropic shipped Claude Opus 4.7 on April 16, and its official migration guide describes the core behavioral change bluntly: 4.7 “takes the instructions literally” and “will not silently generalize.” Huryn argues that many “regression” reports are really friction from workflows that relied on the model to fill in gaps.
Boris Cherny, Claude Code lead at Anthropic, also echoed that adjustment period on release day: “It took a few days for me to learn how to work with it effectively.”
Literal-by-default behavior, with uneven tradeoffs
Huryn characterizes Opus 4.7 as less willing to “guess” when instructions are unclear. That shift has upside for work that benefits from precision—Huryn cites wins in coding, creative writing, and structured tasks—but also introduces losses in areas where ambiguity or iterative clarification was previously tolerated, including vague prompts, multi-turn instruction following, and long-context retrieval.
The thread’s throughline is that the disagreement across Reddit, Arena results, and Anthropic’s own guidance comes down to measuring different workflows rather than a single, universal “better/worse” outcome.
“Intent” as the main lever
Huryn’s headline recommendation is to foreground intent rather than piling on more rules or longer prompts. He splits intent into two layers:
- Strategic context: durable information like what’s being built, who it’s for, constraints, and what “good” looks like—meant to live in
CLAUDE.mdso it can load each session without being retyped. - Per-task intent: the specific request for the current turn, still written each time.
Huryn links this to a broader “Intent Engineering Framework” post and cites Andrej Karpathy’s view that success criteria can outperform imperative instructions for agentic workflows.
He also points to what he calls convergence between vendors: Anthropic moving Opus 4.7 toward more literal instruction following, and OpenAI’s December 2025 Model Spec emphasizing underlying intent.
The 10 Opus 4.7 “migration moves”
Huryn’s thread then lists ten concrete adjustments meant to preserve older 4.6 prompt performance on 4.7. Highlights include:
Put durable context in CLAUDE.md
Front-load strategic context so each new session begins with the same grounding, while pushing less-frequently-needed details into separate files (Huryn mentions strategy.md as an example).
Use the new “xhigh” effort level, and change effort mid-task
Huryn flags a new effort level—extra high (xhigh)—positioned between high and max, and says Anthropic recommends it for coding and agentic work. He also recommends switching effort per call: max only for the hard subproblem, then dropping back down.
Regression-test prompts and budgets
Huryn notes a new tokenizer and estimates 1.0 to 1.35× more tokens per input, meaning older 4.6 workflows can cost more on 4.7 “before you’ve changed a line.” He adds that Anthropic raised rate limits alongside the 4.7 launch, and suggests evaluating “cost per correct output token” with product-specific testing.
Ask in batches, not across multiple turns
One of the more workflow-specific suggestions: stop drip-feeding clarifications. Huryn says multi-turn clarification that worked on 4.6 can create extra overhead on 4.7 because each turn stacks new reasoning on top of earlier literal interpretations.
Show desired outputs instead of stacking “don’t” rules
Huryn argues positive examples land better than negative constraints, citing Anthropic’s guidance that “Like this:” with short examples works, while “Don’t do this:” tends to fail and consume tokens.
Remove progress “scaffolding”
He recommends deleting older prompting patterns like “Explain your plan, then execute” or periodic summary requirements, saying Opus 4.7 already produces higher-quality progress updates during long agentic traces.
Explicitly request fan-out and subagents
Huryn says Opus 4.7 “spawns fewer subagents by default” and makes fewer tool calls per task, meaning parallel exploration has to be requested directly.
Review plans rather than diffs when changes span multiple files
Huryn distinguishes plan review from diff review, pointing to Claude Code CLI plan mode and /ultraplan. The claim is that, with more literal instruction following, intent drift caught in a short plan review can prevent larger downstream misreads in code changes.
“Adaptive thinking” replaces fixed budgets
Finally, Huryn says fixed thinking budgets are gone in Opus 4.7: use thinking: { type: "adaptive" } alongside effort parameters, and update old API calls that still rely on budget_tokens (which he says will return HTTP 400). He links Anthropic’s documentation at adaptive-thinking.md.
Source: Paweł Huryn on X
