An answer is only useful if you know how far to trust it. Language models do not report reliable confidence on their own, and a confidence number a model writes in its answer is a rough signal, not a probability.
Better signals come from behaviour: ask the same question several times and measure how much the answers disagree, or group the answers by meaning and compute semantic entropy. Disagreement can also show which inputs need a person or more examples. These guides cover adding self-check and confidence fields to prompts, detecting when a model is unsure with semantic entropy, and active prompting, which uses uncertainty to choose the few-shot examples worth annotating.
The library lists all guides with filters for topic, level, category and reading time.