Few-Shot Prompt Template

[TASK IN ONE SENTENCE, LEADING WITH A VERB: Classify / Extract / Rewrite / Score.]

[WHY IT MATTERS AND WHO USES THE OUTPUT: one or two sentences. This improves targeting more than any persona line.]

Rules:
- [THE DECISION RULE THE EXAMPLES CANNOT CONVEY ON THEIR OWN]
- [HOW TO HANDLE THE AMBIGUOUS CASE]
- If the input is insufficient to decide, output [THE EXPLICIT UNCERTAIN VALUE] rather than guessing.

Output format: [EXACT FORMAT. For structured output, give the schema and state that no other text may be returned.]

Examples:

Input: [A CLEAR, TYPICAL CASE]
Output: [THE CORRECT OUTPUT, IN EXACTLY THE FORMAT YOU SPECIFIED]

Input: [A CASE WITH A DIFFERENT LABEL OR SHAPE]
Output: [THE CORRECT OUTPUT]

Input: [A GENUINELY HARD OR AMBIGUOUS CASE, THE KIND THAT GETS THIS WRONG IN PRODUCTION]
Output: [THE CORRECT OUTPUT, DEMONSTRATING THE RULE ABOVE]

Input: [A CASE THAT SHOULD PRODUCE THE UNCERTAIN OR EMPTY RESULT]
Output: [THE UNCERTAIN OR EMPTY RESULT]

Now do the same for this input.

Input: [THE REAL INPUT]
Output:

---
Rules for filling this in:

1. Start with ONE example. Add more only when outputs are actually inconsistent. More examples cost tokens and can lock the model into surface patterns you did not intend.
2. Balance the labels. A set of examples where four of five share a label biases the model toward that label regardless of the input.
3. Vary the order between test runs. Example ordering alone can swing accuracy substantially, with the last example exerting the most pull, so an accuracy number from a single ordering is not a measurement.
4. Every example must match your stated rules exactly. Modern models copy the details of examples more faithfully than they follow instructions, so a single example that contradicts a rule silently overrides it.
5. Include the edge cases you actually see in production. Examples made of clean, obvious inputs teach the model nothing about the cases you are trying to fix.
6. On reasoning models, try zero-shot first. Examples often help less there, and sometimes hurt.
7. Format the examples exactly as you want the output formatted, including whitespace and punctuation. That is what the model copies.