When the model feels off, check the plumbing

Anyone who works with these models long enough learns the feeling. The output isn't wrong, exactly. It's just off. The voice has shifted. It's reaching for an idea you'd already ruled out. The vocabulary belongs to a different project. You can't point to a single broken line, but something has drifted.

The temptation is to push back into the prompt and steer harder. That almost never works. The uncanny feeling is signal, not noise, and the cause is almost always upstream of the conversation.

Check the plumbing first

Before regenerating, look at what the model is actually being fed. Is auto-memory turned on for this project, and is it pulling in stale notes you forgot were there? Is there a project-level instructions file loaded — and is it the right one? Are section-level voice rules meant to be imported via references in the system prompt, and are they actually resolving? Did the shape of this task get borrowed from another project, dragging that project's defaults along with it?

Models are faithful channels. If they sound like a different product, it's usually because they're reading from a different product's rules.

A concrete example

A while back the marketing copy on a build kept coming out in an editorial register — measured, magazine-like, longer sentences, more hedging — when the brand was meant to be plain and direct. We rewrote it twice. It snapped back both times.

The cause wasn't the model and it wasn't the prompt. The page scaffolding had been copied from a different project, and one of the import lines in the system instructions — the line that pulled in this project's voice rules — hadn't been carried across. The model was diligently channelling the previous project's voice because that was the only voice it had been given. Once the import was added, the drift stopped on the next generation. No prompt engineering required.

The discipline

When the output feels off, spend a turn on diagnosis before spending five turns on rewriting. Inspect the context the model is being handed — memory, instructions, imports, the lot — and verify that the right inputs are loaded before you blame the output. One diagnostic minute beats an afternoon of polished wrong answers.

The feeling that something is off is the cheapest signal you get. Trust it, and look upstream.