repository·main·Indexed 22 days ago
https://github.com/microsoft/promptyA markdown-based file format for LLM prompts that allows developers to write prompts once and run them across different environments, including Python, TypeScript, and .NET. It uses .prompty files consisting of YAML frontmatter for configuration and a markdown body for prompt templates supporting Jinja2 and Mustache. Prompty provides SDKs for various providers such as OpenAI, Anthropic, and Microsoft Foundry, as well as a VS Code extension for creating, previewing, and debugging prompts.
Prompty is a Microsoft project that provides a specialized asset class and format for LLM (Large Language Model) prompts. It is designed to improve observability, understandability, and portability for developers.
Prompty aims to accelerate the developer inner loop of prompt engineering and prompt source management across different languages and platforms, helping developers build, test, and deploy generative AI applications more efficiently.
Prompty is an asset class and file format for Large Language Model (LLM) prompts. It is designed to improve observability, understandability, and portability during the prompt engineering process, specifically aiming to accelerate the developer inner loop.
Key features include:
.prompty support: Language support and syntax highlighting for the prompt format..tracy files.The Prompty VS Code extension is the primary tool for authoring, running, and debugging .prompty files. It features a built-in TypeScript runtime, meaning no Python installation is required to use it.
Key capabilities include:
.tracy file format with a React-based trace viewer.Prompty is a markdown-based file format for LLM prompts. It uses a single .prompty file to pair YAML frontmatter with a markdown prompt body, making prompts portable and executable across Python, TypeScript, and VS Code.
In a .prompty file, you can define:
The Prompty runtime provides a set of public APIs for managing the lifecycle of a prompt. These APIs allow you to load, prepare, run, and extend prompt execution. The core functional areas include:
.prompty file to execution (load, render, parse, prepare, run, invoke, process).turn), tool dispatching, and agent-mode error handling.connection.kind: reference.outputs, StructuredResult, and cast), and tracing operations.For specific data shapes, frontmatter fields, and model types, refer to the Schema Reference.
The Prompty schema defines the structure for .prompty files, runtime configurations, tool definitions, and message shapes. The schema is the source of truth for how prompts are structured and how the runtime interacts with LLM providers.
Key components include:
Model and Template structures.ApiKeyConnection, OAuthConnection, or FoundryConnection).FunctionTool, McpTool, OpenApiTool).Message and ContentPart objects.TurnOptions (e.g., maxIterations, contextBudget).Prompty supports multiple language runtimes that all share the same .prompty file format. You can choose a runtime based on your existing stack. While the .prompty files are portable, the installation methods, API styles, and provider package names vary by language.
| Language | Status | Package |
|---|---|---|
| Python | ✅ Stable (alpha) | prompty on PyPI |
| TypeScript | ✅ Stable (alpha) | @prompty/core on npm |
| C# | ✅ Alpha | Prompty.Core, Prompty.OpenAI, Prompty.Foundry, Prompty.Anthropic on NuGet |
| Rust | ✅ Alpha | prompty, prompty-openai, prompty-foundry, prompty-anthropic on crates.io |
Connections in the Prompty VS Code extension define how the extension communicates with LLM providers (OpenAI, Anthropic, or Microsoft Foundry). They store endpoints, authentication credentials, and provider-specific settings, allowing you to run .prompty files without manually managing environment variables.
| Provider | Auth Method | Model Discovery |
|---|---|---|
| OpenAI | API key | ✅ Lists available models via the OpenAI API |
| Anthropic | API key | ✅ Lists available models via the Anthropic API |
| Microsoft Foundry | Microsoft Entra ID / API key | ✅ Lists deployed models from your Foundry project |
When running a .prompty file, the extension selects a connection using this priority:
model.connection: Explicitly specified connection name in the .prompty file.isDefault for the provider specified in model.provider.Tools extend LLM capabilities beyond text generation. They are defined in the .prompty file's frontmatter under the tools: key. When a prompt executes, the runtime passes these definitions to the LLM. If the model requests a tool, the agent loop handles the execution and feeds the result back into the conversation.
Every tool requires a name, a kind (which determines its type), and an optional description.
tools:
- name: get_weather
kind: function
description: Get current weather for a city
parameters:
- name: city
kind: string
required: truePrompty is a markdown file format for LLM prompts. A .prompty file combines structured YAML frontmatter (containing model configuration, inputs, tools, and template config) with a markdown body (containing role markers, instructions, and template variables).
The runtime processes these files through a pipeline: loading the file into a typed object, rendering template variables, parsing the markdown into chat messages (system/user/assistant), and executing the request via a provider and model configuration.
A .prompty file is a markdown-based file format used to define LLM prompts. It consists of two main parts:
When parsed, a .prompty file produces a Prompty instance. Runtime loaders can resolve references within the frontmatter such as ${env:VAR} for environment variables and ${file:relative/path} for file references. Note that file references are typically scoped to the .prompty file's directory tree by default.
A .prompty file is a Markdown file containing YAML frontmatter that defines the LLM prompt configuration, including the model, provider, and connection details.
Example .prompty structure:
---
name: greeting
model:
id: gpt-4o-mini
provider: openai
connection:
kind: key
endpoint: ${env:OPENAI_BASE_URL}
apiKey: ${env:OPENAI_API_KEY}
---
system:
You are a helpful assistant.
user:
Hello, my name is {{name}}. {{question}}