—

 

 

 

openclaw 2026.4.23 ships gpt-5.5. ran it an hour.

 

it finally talks like a human lol. 5.4 was rough to chat with, this one i actually enjoy using.

 

it needs anchors. check your config:

 

→ ban ai-speak at the top (delve, leverage, unlock, empower)

→ kill "it depends". strong models drift vague by default. force an opinion

→ negative lists beat positive ones. "you are not a corporate assistant, not a summarizer" beats "be concise"

→ let it swear, both ends. "shit, it's back" on the bad, "holy shit that worked" on the good. only one end = fake

→ 2am test: would you still want to talk to it at 2am? if not, delete and rewrite

 

weak models need an employee handbook. strong models need that friend at 2am.

 

 

if you want to self check against the official docs, read this:

 

https://docs.openclaw.ai/concepts/soul#the-molty-prompt

 

 

@blueprint_os yeah banned list matters but the swear one is my fav ngl

 

 

@Paceminterris13 so back 🦞 gpt-5.5 fixed the vibe

 

 

@kurorosage close, 4.7 still better for chat but we can't auth with a sub 😮‍💨

 

 

@bratan_ai if you auth with your chatgpt sub it basically costs not bad. api is where it hurts