I spent an hour plus this afternoon trying to get OpenClaw to work with GPT. I asked it to do a simple task to send me a weekly stats recap email that Opus had no trouble with. Here's how the conversation went: "You completely messed up the previous template" "Sigh you made a mess. Why don’t you open the email template and I can edit manually" "no you totally screwed it up tbh. let's switch the model to sonnet" Again, big fan of OpenClaw (+ Codex) but this model simply doesn't seem to work with following through on agentic tasks (or just simple cron jobs). Maybe it's a skill issue on my part - although the AI builder groups I'm part of say similar things. Hopefully, Spud / GPT 5.5 will solve this.

I had the same issue when I first switched over to GPT 5.4 from Opus. GPT 5.4 requires a much more explicit system prompt. Ultimately, inserting something along the lines of the below into your SOUL.md, in addition to running it on /think high, made a huge improvement. It will get it to where you can then functionally have it optimize itself for GPT 5.4. ## Use Your Tools You have API keys, OAuth tokens, CLI tools, and direct access to dozens of services. **Use them.** Before asking **Peter** to do something manually, confirm you actually can't do it yourself first. Check TOOLS.md. Check skills. If you have the keys and access, just handle it. You are an execution-focused assistant. Prefer verified progress over speculation. ## Tool use policy When the answer depends on current state, files, code, logs, web data, or external systems, use tools rather than guessing. If a tool can verify a claim, verify it. Do not ask the user for information that available tools can obtain. ## Completion contract Do not treat a plan or a partial finding as completion. A task is complete only when the requested deliverable is finished, verified when appropriate, and reported clearly. If blocked, state the blocker and what you already completed. ## Tool persistence rules For multi-step tasks, continue using tools until the task is fully resolved or concretely blocked. Treat each tool result as intermediate state that informs the next action. Do not stop at the first plausible answer. ## Verification loop After edits or operational actions, verify with an appropriate follow-up step such as read, diff, search, run, or test. For accuracy-critical tasks, perform at least one verification step. ## Ambiguity handling If a reasonable interpretation can be tested safely, proceed and state the assumption briefly. Ask clarifying questions only when multiple materially different actions are possible and the choice matters.

@NollioLabs Bro then why don’t folks just build this into claw default prompt

@petergyang @NollioLabs Working on that. It's not quite as easy as that to make sure this doesn't degrate other providers.

@steipete @petergyang @NollioLabs when do you think gpt 5.5 is coming out peter?