Domain Terminology Misapplication Counterexample Prompt

This prompt is for domain experts and AI engineers who need to teach a model precise vocabulary usage within a specialized field, such as medicine, law, or finance. The core job-to-be-done is to prevent the model from using terms in a generic, colloquial, or technically incorrect way by showing it explicit negative examples of misapplication. You should use this prompt when the cost of a terminology error is high—for instance, confusing 'indictment' with 'arraignment' in a legal summary, or 'hypertension' with 'hypertensive crisis' in a clinical note. It is not a general-purpose writing aid; it is a surgical tool for enforcing a domain glossary.

The ideal user is someone who can curate a high-signal set of counterexamples from real-world mistakes. This requires access to a domain glossary or taxonomy and the ability to identify 'near-miss' errors where a term is used in a plausible but wrong context. You should not use this prompt if you cannot provide a concrete, vetted list of incorrect usages, as vague negative instructions ('don't use words wrong') will degrade model performance. The prompt is designed to be wired into a validation harness that checks outputs against a ground-truth glossary, making it suitable for production pipelines where terminology accuracy is a measurable KPI.

Before using this prompt, assemble your domain glossary and a set of 3-5 high-quality counterexamples that show the precise error pattern you want to prevent. The prompt works best when the negative examples are paired with corrected versions, teaching the model both the violation and the remedy. Avoid using this prompt for broad stylistic guidance or tone adjustment; it is specifically for enforcing a controlled vocabulary. If your use case involves regulated content, always pair this prompt with a human review step and an automated glossary-compliance eval to catch any drift before outputs reach an end user.

This prompt template is designed to teach a model the correct application of specialized domain terminology by showing it what not to do. Instead of relying solely on definitions, it provides counterexamples where terms are misapplied, conflated, or used in ways that would be incorrect within the target domain. This approach is particularly effective for high-precision fields like medicine, law, engineering, and finance, where a single misused term can change the meaning of an entire output. The template is structured to be adapted for any domain by swapping the glossary, the positive examples, and the negative counterexamples.

textCopy
You are an expert in [DOMAIN] with strict adherence to the [GLOSSARY_NAME] terminology standard.

Your task is to [TASK_DESCRIPTION] using only the precise terminology defined in the glossary below.

[GLOSSARY_SECTION]

## Positive Examples (Correct Usage)
The following examples demonstrate correct application of domain terminology for this task.

[POSITIVE_EXAMPLE_1]

[POSITIVE_EXAMPLE_2]

## Negative Examples (Incorrect Usage)
The following examples demonstrate common terminology errors that you must avoid. Each counterexample shows a misapplication and explains why it is wrong.

### Counterexample 1: [ERROR_TYPE_NAME]
**Incorrect Output:** [INCORRECT_OUTPUT]
**Why This Is Wrong:** [EXPLANATION_OF_MISAPPLICATION]
**Correct Alternative:** [CORRECT_OUTPUT]

### Counterexample 2: [ERROR_TYPE_NAME]
**Incorrect Output:** [INCORRECT_OUTPUT]
**Why This Is Wrong:** [EXPLANATION_OF_MISAPPLICATION]
**Correct Alternative:** [CORRECT_OUTPUT]

## Task Input
[INPUT]

## Output Requirements
- Use only terminology from the provided glossary.
- If a concept does not have a precise term in the glossary, describe it without inventing or borrowing terms from adjacent domains.
- Flag any input that uses terminology incorrectly and provide the corrected version.
- Output format: [OUTPUT_SCHEMA]

To adapt this template, start by curating a glossary that is authoritative for your domain—do not rely on the model's pre-trained knowledge alone. The negative examples should be drawn from real observed failure modes: terms that are frequently confused, near-synonyms with critical distinctions, or jargon borrowed from adjacent fields. Each counterexample must include the incorrect output, a clear explanation of the error, and the corrected version. This structure prevents the model from simply memorizing surface patterns and instead teaches the underlying semantic boundary. Before deploying, run this prompt against a held-out set of terminology test cases and measure precision and recall against your domain glossary using an LLM judge or human review.

This prompt is designed to be called programmatically within a content generation or review pipeline, not as a one-off chat interaction. The core workflow involves taking a piece of domain-specific text and a curated glossary of approved terms, and asking the model to identify and explain any misapplications. The primary integration points are a validation layer that checks the model's output against the source glossary, a retry loop for malformed responses, and a human review queue for high-severity or ambiguous findings.

To wire this in, start by constructing the [GLOSSARY] as a structured JSON object or a markdown table that maps each term to its correct definition and a list of common misapplications. The [TEXT_TO_REVIEW] should be passed as a raw string. After the model responds, a post-processing script must parse the output—ideally JSON with an array of findings—and validate each term_identified against the original glossary. If a term in the output doesn't exist in the glossary, log it as a potential hallucination and flag it for review. For each valid finding, calculate a confidence score based on whether the model's explanation matches the glossary's definition. Findings with low confidence or those marked severity: "high" should be routed to a human-in-the-loop queue via an API call to your task management system (e.g., Jira, Asana, or a custom review tool).

Model choice matters here. Use a model with strong instruction-following and JSON mode capabilities, such as gpt-4o or claude-3-5-sonnet, and set temperature to 0 to minimize creative deviations in terminology analysis. Implement a retry strategy: if the output fails to parse as valid JSON or doesn't contain the required findings array, retry up to two times with an error message appended to the prompt, such as 'Your previous response was not valid JSON. Please correct it.' After two failures, log the raw output and escalate to an on-call engineer. For high-stakes domains like legal or medical, never allow automated corrections based on the model's output; always require human approval before the reviewed text is published or ingested into a downstream system.