Claude Opus 4.7 Prompting
🚨 Claude Opus 4.7 is out. It's state-of-the-art on GDPval-AA (a third-party eval of real economically-valuable knowledge work), and it's the new default for the hardest tasks. Anthropic buried something important in their release notes: "Prompts written for earlier models can sometimes now produce unexpected results: where previous models interpreted instructions loosely or skipped parts entirely, Opus 4.7 takes the instructions literally. Users should re-tune their prompts and harnesses accordingly." Translation: prompts that worked on Opus 4.6, Sonnet 4.6, or any prior Claude model might quietly degrade on 4.7. Not because the model got worse. Because it got more literal. And this applies doubly to prompts written for other models. Your GPT-5.4 prompts with ::: markers and output contracts. Your Gemini prompts with data-at-top and negative constraints at the bottom. Those patterns were optimized for those models. On Opus 4.7, they produce friction. A few examples of what changes: If your prompt has two competing instructions (like "be concise" and "give a comprehensive analysis"), 4.6 smoothed that over. 4.7 picks one literally. "Challenge my assumptions" was useful on 4.6 and standard on Grok. 4.7 pushes back natively. Leaving the line in is clutter. "Be thorough" and "don't skip steps" were hedges against older models cutting corners. 4.7 is proactive by default. More clutter. Dense screenshots used to need downsampling. 4.7 handles images up to 2,576px on the long edge (3× the old resolution). I built a skill to fix all of this. 👾Join Monday's masterclass to learn how: https://links.convertwave.ai/widget/form/tElay6We4IZgzcXxsbuy?contact_id=U8htJWSeFEIBtYJWIYQf&fbclid=PAT01DUAROb7xleHRuA2FlbQIxMABzcnRjBmFwcF9pZA81NjcwNjczNDMzNTI0MjcAAacah6qLq0J1Zb1Ex_FkhWZ55nUHosab749SbfYtyGIlpUjYd3dsSFQou-COlg_aem_AJ5J2V65cSh3yLZ1KQttSw