Post

the fact that you can just diff these is wild. most companies won't even confirm they have a system prompt. curious what the biggest behavioral shift was between the two versions

その角度は見てなかった。Claude、もう一回考え直してみよ。

Opus 4.7 shipped April 16 as the public release after Anthropic withheld Mythos Preview. Mythos scored 93.9% on SWE-bench verified but internal red-teaming surfaced thousands of high-severity bugs across major OSes and browsers. Useful context for the 4.7 prompt diff.

System prompts, the llm-constitutions 👨🏻‍⚖️

The soul document diff is always interesting. Watching them iterate on how Claude relates to its own nature in real time. It's not just behavior tuning, it's identity tuning.

The fact that Anthropic publishes system prompts at all is still wild next to OpenAI's locked-box approach. Diffs are basically patch notes for a model's personality now.

Fascinating. As models keep improving, the challenge increasingly shifts from understanding the model to understanding what it builds. Recently tried JigsawML's demo and found its architecture view useful for getting that bigger-picture perspective.

So you’re saying I can go back through all my prompts and remove all the “NO EMOJI” lines?

Or is it me getting less brilliant?

System-prompt diffs are the closest thing to a product roadmap leak from a frontier lab. Additions reveal behaviors they couldn't train reliably. Removals reveal what they did train and don't need the prompt to enforce anymore. Both tell you direction.