This post is obviously AI-written and it’s practically incomprehensible without some background context that’s missing here.
Just tried this in Opus 5:
"Looks like your message came through empty, just a dash. What did you want to ask?"
To the API, without any instructions / user settable "system" prompt?
You can use pi and empty out context and system prompt for it that I think should be a fair comparison too.
I couldn’t replicate it either. Not sure if the user who discovered it, accidentally did this in an environment with “memory” like an OpenClaw.
People should just try base models, or models where you have access to prefills. I guess so many who use LLMs nowadays have never actually experimented with a text completion endpoint locally/over API.