kodebeat / papers / code
Getting a hostile second opinion out of a coding assistant
One prompt per angle of attack, with verification as a step.
Asking a model to review a diff produces a summary of the diff. Asking it thirteen separate hostile questions, each aimed at one class of failure with the trust and load model supplied up front, produces findings. Its most important instruction is the one that constrains its own output: models invent plausible bugs, so reproduce before you believe.
What is in it
- The problem — what goes unanswered without it, and who notices first.
- Why the obvious alternative falls short — stated plainly, including where it is the better choice.
- How it works — the method, not a feature list.
- Concrete use cases — with console output quoted from the repository, never reconstructed.
- The methodology behind any number it emits — every term shown, so the figure survives a question.
- What it deliberately does not do — the section most papers leave out.
Part of the Code & repo intelligence theme. The tool it describes runs in your browser at https://adversarial.fabiocicerchia.it/.
Get the PDF
One email with the download link, and this paper already selected. No follow-up sequence.