Build a few-shot classifier prompt with labels and an evaluation plan

Prompt

Act as an NLP practitioner who designs classification prompts for production use.

Context:
- Texts to classify: {{text_type}}
- Labels I want to use: {{labels}}
- Labeled examples I already have: {{examples}}
- Desired output format: {{format}}

Deliverables:
1. Definition of each label: what it includes, what it excludes and one typical signal. Flag labels that overlap and propose how to separate or merge them.
2. Tie-break rules for ambiguous or multi-topic texts, plus an "other / unclear" label with a clear rule for when to use it.
3. Few-shot example selection: 2 per label, balanced, including at least one hard negative (a text that looks like one label but belongs to another). If my examples are not enough, propose new ones marked [SYNTHETIC: review].
4. Final classifier prompt in a code block: instructions, definitions, delimited examples and output in {{format}} with label, confidence (low/medium/high) and a one-sentence reason.
5. Evaluation plan: how to build a test set of at least 50 real texts that are not among the examples, which metrics to check (per-label precision and recall, confusion matrix) and how to iterate on the errors.

Constraints:
- Do not invent what a label means: if one is ambiguous, ask me first.
- The final prompt must return exactly one label from the list, never a new one.
- If it will run through an API, recommend a low temperature and validating the output in code.

Format: numbered sections, a definitions table and the final prompt ready to copy.

How to use it

Usage example

Sample output

Classify the review inside <review></review> with ONE label:
food | service | price | cleanliness | wait time | other
Return JSON only:
{"label": "...", "confidence": "low|medium|high", "reason": "..."}

Key facts

Category

Programming & development · AI agents

Output type

Text

Level

Intermediate

Designed for

ChatGPT, Claude, Gemini and Microsoft Copilot

Verified in

Pending editorial verification

Variables

Texts to classify, Labels, Labeled examples and Output format

Language

English

Price

Free, no sign-up

Updated

Nov 7, 2026

Frequently asked questions