PROMPT:
├─ Classification question
├─ Criteria for category 1 (relevant): list of signals
├─ Criteria for category 2 (partial): list of signals  
├─ Criteria for category 3 (irrelevant): catch-all category
├─ [Optional] 2-3 examples with explanations (few-shot)
└─ Output format: strict instruction to return one word

OUTPUT: One word from the three categories

Runs in a single request.

Task: You're filtering 200 AI papers for a subscriber newsletter. You only want the ones with practical prompting techniques, not theory or infrastructure.

Prompt:

How useful is this paper for a reader who works with ChatGPT/Claude in a chat window and wants to learn new prompting techniques?

A **Relevant** paper:
- Describes a ready-to-use prompting technique that works in chat (Chain-of-Thought, Few-Shot, Role Prompting, etc.)
- Provides prompt templates or examples you can copy

A **Partially relevant** paper:
- Mentions prompting, but focuses on theory or benchmarks
- Explains how LLMs work in a way that aids understanding, but offers no ready-made techniques

An **Irrelevant** paper:
- Is about fine-tuning, model training, infrastructure, API development
- Is about applying LLMs without a focus on prompting

**FORMAT:** Answer with one word only: Relevant, Partial, Irrelevant

---

[Paste the paper's title + abstract]

Result:

The model returns a single word. You filter 200 papers in 2-3 minutes: copy the prompt, paste the text, get the category. Relevant ones go into the newsletter, partial ones get parked for later, irrelevant ones get skipped.

In "yes/no" mode, an LLM tends to flatten the gray zone: a paper might be partly useful, but the model shoves it into one extreme or the other. Three categories create room for nuance: if a paper mentions the topic but offers no practical value, it's "partial," not "no."

Detailed criteria act as a checklist inside the model's head. Without criteria it relies on a generic semantic match ("the words look similar, so it's relevant"). With criteria it checks specific signals: is there a cause-and-effect link? Are there practical examples? That beats "looks on-topic" by a mile.

Few-shot examples calibrate the category boundaries. The model sees: this paper mentions both terms but draws no link, so it's "partial." This one explains a mechanism, so it's "relevant." Without examples the model invents the boundaries itself; with examples it copies the pattern.

A strict output format ("one word only") cuts off explanations and hedging. The model can't write "probably relevant, but...": it has to choose. That speeds up processing and makes the result machine-parseable.

Control Levers

• Number of categories → drop to two (relevant/irrelevant) for simple tasks or expand to four or five for finer sorting
• Criteria detail → the more detailed, the more stable, but the longer the prompt; you can simplify to 1-2 signals per category
• Few-shot examples → 2-3 examples sharply improve accuracy; beyond 5 the gains fade while token costs keep climbing
• Output format → you can let the model explain its choice after the category if you need transparency (but it'll be slower and harder to parse)

How well does {item} fit the criterion: {your criterion}?

**{Category 1}** {item}:
- {Signal 1}
- {Signal 2}

**{Category 2}** {item}:
- {Signal 1}
- {Signal 2}

**{Category 3}** {item}:
- {Signal 1} OR does not meet the criteria above

**FORMAT:** Answer with only one of three words: {Category 1}, {Category 2}, {Category 3}

---

[Paste the item's content: paper text, candidate profile, customer review, etc.]

What to fill in: - {item}: what you're classifying: a paper, a résumé, a review, a comment - {criterion}: what you're sorting by: usefulness for your audience, fit for a job opening, sentiment - {Category 1/2/3}: category names: relevant/partial/irrelevant; good fit/maybe/no fit; positive/neutral/negative - {Signal}: concrete indicators: topic mentioned, numbers present, cause-and-effect link, work experience, emotional tone

To boost accuracy (few-shot):

Add 2-3 examples after the criteria, before the format instruction:

**Example 1:**
[Item text]
→ {Category 1}

**Example 2:**
[Item text]
→ {Category 2}

**Example 3:**
[Item text]
→ {Category 3}

The examples should be edge cases, where the difference between categories isn't obvious. That's what tunes the model to nuance.

⚠️ Random artifacts: The model can latch onto individual words instead of meaning. The study showed that removing a single term sometimes changed the category even though the meaning stayed the same. The more complex the criteria, the higher the risk. Fix: test manually on 5-10 items and check for stability.

⚠️ Subjective criteria: If the line between categories is fuzzy even to you ("creativity," "growth potential"), the model will give unstable results. Three categories won't save you if the criteria themselves are vague. Fix: operationalize your criteria: instead of "creativity," use "uses metaphors or unexpected comparisons."

⚠️ Context length: With long texts (>2000 words) the model may skim and miss details in the middle. Fix: for large documents, summarize first, then classify the summary.

Small Language Models Can Use Nuanced Reasoning For Health Science Research Classification: A Microbial-Oncogenesis Case Study

Authors: Muhammed Muaaz Dawood, Mohammad Zaid Moonsamy, Kaela Kokkas, Hairong Wang, Robert F. Breiman, Richard Klein, Emmanuel K. Sekyi, Bruce A. Bassett

University of the Witwatersrand (South Africa), OncoVectra (UK)