Reading A Postmortem Out Of Raw Logs With One Prompt

Pasting 4,000 log lines and asking "what went wrong" gets you a confident wrong answer. The model will find a pattern and commit to it.

The prompt that keeps it honest:

"Here are the logs from 14:02 to 14:41 UTC across three services.

Step 1: build a timeline of distinct events. Do not interpret yet.
Step 2: for each event, mark whether it is plausibly a cause, plausibly an effect, or unclear.
Step 3: list the candidate root causes with the specific log lines supporting each.
Step 4: say explicitly what you would need to check to distinguish between them.

If the logs do not determine the cause, say so. Do not pick the most likely one."

Step 4 is what makes this useful. It turns the output into a list of things to go look at, rather than a conclusion you then have to disprove.