Example-led calibration

Using selected examples and counterexamples to clarify a task's intended output distribution, criteria, or boundary cases.

Example-led calibration is steering an AI's output by showing it a few samples at the difficulty, tone, or format you want, so it copies the standard you demonstrated instead of guessing from a description. You give the agent two or three finished examples and ask it to match them, rather than spelling out the standard in the abstract. Say you want practice questions for a study guide and your first batch comes back too easy. Instead of repeating "make them harder," you paste two questions at the exact level you have in mind and ask for ten more like those. The samples pin down what "harder" means far more precisely than an adjective can, and the agent calibrates its next batch to the demonstrated bar.

Builder example

Examples can specify properties that adjectives leave vague, but they can also cause imitation of irrelevant surface features or expose private data. Representative coverage, labels explaining why an example passes, and held-out evaluation matter more than a fixed number of samples.

Common confusion: Examples are evidence about the target, not the target itself. A narrow set may encode accidental length, topic, or demographic patterns that should not generalize.