I’ve also had trouble replicating it on claude.ai. But I find on openrouter, if you disable reasoning and use the default sysprompt it works almost every time.
I was unable to replicate this in Claude.AI with 30+ tries with memory off. I also tried incognito, no help. I was able to replicate by using API via platform.claude.com. I used Opus 5 and Fable with low reasoning.
I’m unable to replicate. I’m on a Max plan, I used Opus 5 with Low thinking, and selected “Quick answer” to bypass thinking. What am I doing wrong? https://claude.ai/share/b7fa8848-c066-4c09-ae35-f6134c0a079f.
Use—instead of —
This should be higher up, it does seem to actually work when you do that. I’m able to replicate now.
Thanks! I can now replicate with a prompt of
on Opus 5 with Low or Medium reasoning, without clicking “Quick answer”. About 70% of the time, it exhibits the base-model-like behavior: https://claude.ai/share/9c46d7de-b3d1-477b-ae28-b6eb8af904e9.
I’ve also had trouble replicating it on claude.ai. But I find on openrouter, if you disable reasoning and use the default sysprompt it works almost every time.
This version of the prompt seems to work pretty reliably
https://claude.ai/share/b5cb104a-69f8-4de5-a354-944a077c452b
Strange
I was unable to replicate this in Claude.AI with 30+ tries with memory off. I also tried incognito, no help. I was able to replicate by using API via platform.claude.com. I used Opus 5 and Fable with low reasoning.
Patched on claude.ai as well for me; it worked with user instructions before.
Also failing to replicate, same setup and no personal instructions.