Claude Opus 4.7 started following prompts much more literally, and every model since has built on that. Here’s what changed, in Anthropic’s own words, and how to prompt Claude today.
Anthropic released Claude Opus 4.7 in April 2026 and told people to re-tune their prompts. If your results have felt shorter, more literal or stuck at the bare minimum, these are the fixes.
Quick update: the newest models are now Opus 5.5, Fable 5.1 and Sonnet 5.5. Most of this still applies, and I’ve noted where things changed.
01
It takes you literally
Anthropic, Opus 4.7 launch
"Opus 4.7 takes the instructions literally."
Claude won’t guess at extra things you didn’t ask for. Anthropic’s prompting guide says it "does not infer requests you didn’t make," so spell out every step and say how far an instruction should reach.
Instead of: "Clean up this draft."
Try: "Fix typos, shorten any sentence over 25 words and cut anything repetitive. Keep the structure. Apply this to every section."
02
Ask for the length you want
Opus 4.7 and 4.8 matched answer length to how hard the question looked, so simple asks got very short answers. The newest Opus models go the other way. Anthropic says Opus 5’s answers "run longer than prior models’." Either way, the fix is the same: say exactly how long you want it.
Instead of: "Summarize this."
Try: "Give me a full breakdown in 3 to 4 paragraphs: the main argument, the evidence and the strongest counterargument."
03
Drop the ALL CAPS
Anthropic prompting guide
"The fix is to soften any aggressive language."
Anthropic flagged this when Opus 4.5 and 4.6 started following prompts more closely: words like "CRITICAL" and "YOU MUST" can make Claude overreact. Just talk to it normally.
Instead of: "CRITICAL: You MUST use the calculator for ANY math."
Try: "Use the calculator for math problems."
04
Ask it to go beyond the basics
Anthropic prompting guide
"If you want Claude to do more than the task strictly requires, explicitly request it rather than relying on the model to infer this from vague prompts."
For creative work, Anthropic’s own example adds: "Go beyond the basics to create a fully-featured implementation."
Instead of: "Design a landing page for my app."
Try: "Design a landing page for my app. Go beyond the basics. Include a hero section, social proof, features, pricing, an FAQ and a final call to action."
05
Thinking moved to the Effort setting
On Opus 4.7, Claude’s deeper thinking was off by default in the API, so people added "think carefully" to their prompts. That’s changed. On today’s models, Claude always thinks in the Claude app, and the control is Effort in the model menu next to the send button, with five levels from Low up to Max.
For hard problems, raise the Effort instead of adding "think carefully." For Opus 5.5, Anthropic now suggests removing those lines, because effort is the main control.
None of these are hacks. Claude got better at doing exactly what you say, so the skill now is saying exactly what you want. Want Claude set up right from the start? Here’s my 5-step setup.