prompt-spec

A lightweight, LinkML-aligned prompt and schema management toolkit for clinical NLP & LLM workflows.

prompt-spec provides:

• Structured prompt templates
• LinkML → Pydantic auto-generation for LLM output schemas
• Validation of few-shot examples within prompts
• A shared format for prompt libraries across collaborating groups

This toolkit is intentionally minimal and designed for stability, reuse, and strong typing when integrating with LLM libraries such as pydantic-instructor

---

Project Structure

prompt-spec/
├── cli.py                        ← Main command line interface
├── core/
│   └── prompt_template.py        ← Prompt class utilities
├── engines/
│   ├── __init__.py
│   └── instructor_engine.py      ← Wrapper function for integrating prompts with ontoGPT via Instructor interface
├── output_models/
│   └── condition_model.yaml      ← LinkML schema for defining model outputs - create new endpoint definitions here
├── generated_models/
│   └── (auto-generated pydantic models)
└── prompts/
    └── condition_prompt.yaml     ← This is where the actual prompt definitions exist

---

Installation

uv pip install -e .

---

Usage

1. Defining LinkML output models

Place all LinkML schemas inside prompt-spec/output_models

Example: condition_model.yaml

name: ConditionList
id: https://example.org/condition_model
prefixes:
  linkml: https://w3id.org/linkml/
default_range: string

classes:
  Condition:
    attributes:
      label: string
      verbatim_name: string
      codable_name: string
      who_diagnosed: string
      is_negated: boolean

  ConditionList:
    attributes:
      conditions:
        multivalued: true
        range: Condition

2. Generate Pydantic models from LinkML

prompt-spec build-models

This reads all *.yaml schemas in prompt-spec/output_models/ and generates Pydantic classes into prompt-spec/generated_models/.

After running this, you will have files such as:

• prompt-spec/generated_models/ConditionList.py

Alternatively, run for just one class update at a time: prompt-spec generate-pydantic-from-linkml output_models/condition_model.yaml

3. Populate prompt template

prompt-spec create-empty-prompt condition_model prompts/condition_prompt.yaml

This produces:

output_model: condition_model
instruction: "<<< fill in your system prompt here >>>"
examples:
  example_1:
    input: "<<< sample input text here >>>"
    output:
      conditions: []

4. Validate a prompt against its template

prompt-spec validate-prompt condition_prompt.yaml