GPT-5.4 in OpenClaw was unusably lazy — empty tool args, infinite retries, rate limit death spirals.
Fixed it by:
Rewriting SOUL.md with completeness contracts + tool persistence rules
Adding "DO NOT BE LAZY" execution standards to AGENTS.md
Setting thinking to medium
Key insight from OpenAI's own prompt guidance: the model needs explicit completion contracts and verification loops. Without them it satisfies at the first plausible answer.
Still not as reliable as Claude for tool use, but went from 48 errors/session to actually completing tasks.
GPT 5.4 so lame!
Just uesd gpt 5.4 in openclaw one full day, it's so lame!
Lazy, failed complicated tasks many times, not good at muti rounds tasks at all, and not proactive at all and, speaking so AI and sick, omg😅 compared with the claude opus or sonnect is a joke
Just uesd gpt 5.4 in openclaw one full day, it's so lame!
Lazy, failed complicated tasks many times, not good at muti rounds tasks at all, and not proactive at all
and, speaking so AI and sick, omg😅
@DanTheronX Ya this works I know but the cost is so high. 😅I spent $100 one day, so I’m looking for sth like the Claude subscription can be used in the Openclaw. Planed to try codex today