Ground the Agent with Examples

Abstract instructions are easy to misinterpret. Concrete examples are hard to misunderstand. If you want the agent to match a style, format, or standard, show it what matching looks like.

Large language models are pattern matchers trained on examples. When you give them an example, you activate the right region of their training distribution. When you only describe the pattern, you force them to translate your description back into behavior — and translations are lossy.

Bad: "Write the response in a friendly tone."

Better: "Write the response like this: 'Got it — I’ll take care of that for you and let you know when it’s done.'"

Bad: "Format the output as JSON."

Better: "Return JSON exactly like this structure: {"status": "ok", "items": [...]}"

• Input/output pairs. Show a sample input and the desired output.
• Style samples. Paste a paragraph written the way you want.
• Edge cases. Show how to handle null values, errors, or empty results.
• Counter-examples. Show what not to do if the agent has a recurring bad habit.

Put examples near the top of the prompt, right after the role and task. If the task is complex, include a "Examples" section with 2–3 cases before asking for the main output.

Examples add tokens. But a good example usually saves more tokens than it costs by reducing retries and corrections. Treat examples as an investment, not overhead.