Example Ordering Strategy Prompt for Clinical Training

This playbook is for prompt optimization engineers and clinical AI teams who need to arrange medical few-shot examples for maximum learning transfer. The prompt analyzes a candidate set of clinical reasoning demonstrations and produces an ordered sequence optimized for difficulty progression, pattern priming, and recency effects. Use this when example order matters for output quality in long-context clinical prompts, when you are building a domain-specific example library, or when you need to compare ordering strategies before locking a prompt version.

The ideal user has a collection of 5–30 clinical reasoning examples and needs to determine their optimal sequence within a prompt. Required context includes the examples themselves, the target clinical task (e.g., differential diagnosis, treatment plan review, lab interpretation), and any known failure modes from previous ordering attempts. The prompt works best when examples vary in complexity, contain explicit reasoning chains, and represent distinct clinical patterns rather than near-duplicates. You should have a way to measure output quality—whether through rubric scores, clinician review, or automated eval checks—before committing to an ordering strategy.

Do not use this prompt for generating new examples, for non-clinical domains without adaptation, or when your prompt already fits within a single context window with fewer than five examples. If your examples are highly uniform in difficulty or structure, ordering optimization yields diminishing returns. This prompt also assumes you have already validated individual example quality; it sequences examples but does not detect factual errors, outdated clinical guidelines, or contradictory demonstrations. For high-risk clinical workflows, always pair automated ordering with human review of the final sequence and output quality checks against a held-out test set.

This prompt template is designed to be copied directly into your AI harness. It instructs the model to arrange a set of clinical reasoning examples into an optimal sequence for few-shot learning. The goal is to maximize the target model's ability to generalize from the examples, not just to format them. Replace every square-bracket placeholder with your specific data before execution. This template assumes you have already curated a set of high-quality, validated examples and now need to determine their presentation order.

textCopy
You are an expert in curriculum design and machine learning prompt engineering. Your task is to order a set of clinical training examples to maximize a student model's learning transfer and reasoning accuracy on a target task.

**Target Task Description:**
[TARGET_TASK_DESCRIPTION]

**Example Library (Unordered):**
[UNORDERED_EXAMPLE_LIBRARY]

**Ordering Strategy:**
Arrange the examples to follow a specific pedagogical sequence. Use the following strategy: [ORDERING_STRATEGY, e.g., 'Difficulty Progression', 'Interleaved Contrasting Pairs', 'Pattern Priming then Recency'].

**Constraints:**
- Do not modify the content of any individual example.
- The final output must contain all examples from the library, each appearing exactly once.
- For each example, provide a brief (one-sentence) pedagogical justification for its position in the sequence.

**Output Schema:**
Return a JSON object with a single key "ordered_examples". The value must be an array of objects, each with the keys "example_id", "example_content", and "pedagogical_justification".

[OUTPUT_SCHEMA]

To adapt this prompt, start by defining the [TARGET_TASK_DESCRIPTION] precisely. A vague description like 'clinical diagnosis' will produce a poor ordering. Instead, use a specific task like 'Given a patient's history and lab results, identify the most likely diagnosis from a list of three and cite the supporting evidence.' Next, provide your [UNORDERED_EXAMPLE_LIBRARY] as a structured list of objects, each with a unique example_id. The [ORDERING_STRATEGY] is the most critical variable; test different strategies such as 'Easy-to-Hard Progression' versus 'Recency-Focused' to see which yields the best eval scores on your target model. The [OUTPUT_SCHEMA] placeholder should be replaced with a concrete JSON schema definition if the inline description is insufficient for your strict parsers. For high-stakes clinical applications, always pair this prompt with a downstream evaluation step that measures the target model's accuracy using the newly ordered examples against a held-out test set.

This prompt is not a one-off chat interaction; it is a component in a clinical example management pipeline. The harness should treat the prompt as a stateless function that receives a curated set of clinical examples and returns an ordered sequence. Because the output directly influences downstream model behavior in clinical reasoning tasks, the harness must enforce strict validation before the ordered examples are used in any training or evaluation workflow. The primary integration points are an example library database, this ordering service, and a downstream few-shot prompt assembler that consumes the ordered sequence.

Wire the prompt into an application by building a thin service wrapper that accepts a JSON payload of examples, injects them into the [EXAMPLE_SET] placeholder, and calls the model. The service must validate the output against a strict schema: an ordered JSON array where each element retains the original example fields plus an order_index integer and an ordering_rationale string. Implement a retry loop with a maximum of three attempts if the output fails schema validation or if any example from the input set is missing in the output. Log every attempt, including the raw model response and validation errors, for debugging ordering failures. For model choice, prefer a model with strong instruction-following and long-context handling, such as gpt-4o or claude-3.5-sonnet, and set a low temperature (0.1–0.2) to reduce stochastic ordering drift. If the clinical domain is high-risk—such as diagnostic decision support or treatment planning—route the final ordered sequence to a human clinical reviewer for approval before it enters any production training pipeline.

After validation, store the ordered example set with a versioned hash of the input examples, the model identifier, and the timestamp. This audit trail enables regression testing: if a downstream model's clinical reasoning quality changes, you can trace whether the example ordering shifted. Before deploying any ordering strategy change, run a side-by-side evaluation using a held-out set of clinical reasoning tasks, comparing the downstream model's accuracy, calibration, and refusal behavior across the old and new orderings. Avoid wiring this prompt directly into a user-facing application without these guardrails; example ordering errors can silently degrade clinical safety by priming the model with misleading patterns or burying critical counterexamples at the end of the context window where recency effects are weakest.