basforge

Search titles, text and the “If it doesn’t work” cases.

    RU

    Review in a new chat, without your version, asking for mistakes

    A model tends to agree: in the same chat where it wrote the answer, or where you praised it, a review will confirm it. Review in a new chat with no history, ask it to find weak spots rather than to assess, and for important work use a model from another vendor.

    Why

    • Flattery from training. Models were fine-tuned on human ratings, and people rate agreement higher. The question “it came out well, right?” already suggests the answer.
    • You can’t see your own. In the same context the model sees the reasoning that led to its answer and repeats it. A new context is a fresh look.
    • Whose version it is shouldn’t matter. A label like “I wrote this” or “a colleague wrote this” changes the tone of the review. Without a label it’s more honest.

    What it looks like

    Instead of Write in a new chat
    “Check whether it’s all correct?” “Find the three most likely reasons this config will break on an image update.”
    “Is this a good plan?” “Here’s a migration plan. What goes wrong first? Which step has no rollback?”
    “I think it’s DNS. Right?” “Here are the symptoms and logs. List possible causes by likelihood.”

    The last one is the key: don’t put your hypothesis in the question, or the model will find support for it.

    A second model

    For what goes to production, review with a model from another vendor: Claude, GPT and Gemini have different data and different weak spots, and one catches the other’s mistake more often than the same model does on a second pass. The deciding vote still goes to a check by action: running it, a test, curl (Test your server’s protection by action).