promptdown 1.1.6

A package for loading promptdown files, which are a special type of markdown file for defining structured LLM prompts

Promptdown

Promptdown is a Python package that allows you to express structured prompts for language models in a markdown format. It provides a simple and intuitive way to define and manage prompts, making it easier to work with language models in your projects.

Installation

Using PDM

Promptdown can be installed using PDM:

pdm add promptdown

Using pip

Alternatively, you can install Promptdown using pip:

pip install promptdown

Usage

Basic Usage

To use Promptdown, simply create a Promptdown file (.prompt.md) with the following format. You can use either a System Message or a Developer Message (for newer model APIs), but not both:

# My Prompt

## System Message

You are a helpful assistant.

## Conversation

**User:**
Hi, can you help me?

**Assistant:**
Of course! What do you need assistance with?

**User:**
I'm having trouble with my code.

**Assistant:**
I'd be happy to help. What seems to be the problem?

Or alternatively:

# My Prompt

## Developer Message

You are a helpful assistant.

## Conversation

**User:**
Hi, can you help me?

**Assistant:**
Of course! What do you need assistance with?

**User:**
I'm having trouble with my code.

**Assistant:**
I'd be happy to help. What seems to be the problem?

Then, you can parse this file into a StructuredPrompt object using Promptdown:

from promptdown import StructuredPrompt

structured_prompt = StructuredPrompt.from_promptdown_file('path/to/your_prompt_file.prompt.md')
print(structured_prompt)

Please note that:

• The Conversation section can be omitted
• Either a System Message or Developer Message section is required, but not both
• Use Developer Message for newer model APIs (like OpenAI's o1) that expect the "developer" role instead of "system"
• Conversations use a simplified format where roles are marked with bold text (**User:** or **Assistant:**, optionally **Role (Name):** to include a name). Only User and Assistant roles are recognized; other roles are ignored with a warning, and conversation roles cannot be system or developer.
• Conversation message lines are collapsed into a single line with spaces (blank lines are dropped).

Parsing a Prompt from a String

For scenarios where you have the prompt data as a string (perhaps dynamically generated or retrieved from an external source), you can parse it directly:

from promptdown import StructuredPrompt

promptdown_string = """
# My Prompt

## Developer Message

You are a helpful assistant.

## Conversation

**User:**
Hi, can you help me?

**Assistant:**
Of course! What do you need assistance with?

**User:**
I'm having trouble with my code.

**Assistant:**
I'd be happy to help. What seems to be the problem?
"""

structured_prompt = StructuredPrompt.from_promptdown_string(promptdown_string)
print(structured_prompt)

Converting to Chat Completion Messages

The to_chat_completion_messages method converts a StructuredPrompt instance into a list of dictionaries suitable for chat completion API clients. The returned list includes the system or developer message first, followed by the conversation messages. This is useful when you need to send the structured conversation to an API that expects messages in a specific format. Here's an example of how to use this method:

from promptdown import StructuredPrompt

promptdown_string = """
# My Prompt

## Developer Message

You are a helpful assistant.

## Conversation

**User:**
Hi, can you help me?

**Assistant:**
Of course! What do you need assistance with?

**User:**
I'm having trouble with my code.
"""

structured_prompt = StructuredPrompt.from_promptdown_string(promptdown_string)
messages_from_promptdown = structured_prompt.to_chat_completion_messages()

response = client.chat.completions.create(
    model="gpt-4o",
    messages=messages_from_promptdown,
    temperature=0.7,
    max_tokens=300,
)

Converting to OpenAI Responses input

For the OpenAI Responses API, use to_responses_input() to emit messages in the expected format. Any prior system content is mapped to the developer role by default for consistency with newer models; you can disable this via map_system_to_developer=False.

from promptdown import StructuredPrompt

structured_prompt = StructuredPrompt.from_promptdown_string(promptdown_string)
responses_input = structured_prompt.to_responses_input()

# Example with the OpenAI SDK (Responses API)
from openai import OpenAI
client = OpenAI()

result = client.responses.create(
    model="gpt-5",
    input=responses_input,
    reasoning={"effort": "medium"},
)

Notes:

• to_responses_input() outputs a list of messages, each { "role": "<role>", "content": [{"type": "input_text", "text": "..."}] }.
• Content is always non-null; non-strings are coerced with str(...).
• Already-structured input_text parts are passed through; other shapes are coerced to text.
• Current scope focuses on text parts; additional types (images/tools) can be added in the future.

If you have legacy Chat Completions-style messages and want to convert them to Responses input, a convenience converter is available:

from promptdown.converters import convert_chat_messages_to_responses_input

legacy_messages = [
    {"role": "system", "content": "You are helpful."},
    {"role": "user", "content": [{"type": "text", "text": "Hello"}]},
]
responses_messages = convert_chat_messages_to_responses_input(legacy_messages)

Loading Prompts from Package Resources

For applications where prompts are bundled within Python packages, Promptdown can load prompts directly from these resources. This approach is useful for distributing prompts alongside Python libraries or applications:

from promptdown import StructuredPrompt

structured_prompt = StructuredPrompt.from_package_resource('your_package', 'your_prompt_file.prompt.md')
print(structured_prompt)

This method facilitates easy management of prompts within a package, ensuring that they can be versioned, shared, and reused effectively.

Using Template Strings

Promptdown supports the use of template strings within your prompts, allowing for dynamic customization of both system messages and conversation content. This feature is particularly useful when you need to tailor prompts based on specific contexts or user data.

Defining Template Strings

To incorporate template strings in your Promptdown files, use curly braces {variable} around placeholders that you intend to replace dynamically. Here is an example of how to use template strings in a prompt:

# My Prompt

## Developer Message

You are a helpful assistant in {topic}.

## Conversation

**User:**
Hi, can you help me with {topic}?

**Assistant:**
Of course! What specifically do you need help with in {topic}?

**User:**
I'm having trouble understanding {concept}.

**Assistant:**
No problem! Let's dive into {concept} together.

Applying Template Values

Once you have defined a prompt with placeholders, you can replace these placeholders by passing a dictionary of template values to the apply_template_values method. Here's how you can apply template values to your prompt:

from promptdown import StructuredPrompt

# Load your structured prompt from a file or string that contains template placeholders
structured_prompt = StructuredPrompt.from_promptdown_string(promptdown_string)

# Define the template values to apply
template_values = {
    "topic": "Python programming",
    "concept": "decorators"
}

# Apply the template values (returns a new StructuredPrompt)
new_prompt = structured_prompt.apply_template_values(template_values)

# Output the updated prompt
print(new_prompt)

This returns a new prompt where {topic} becomes "Python programming" and {concept} becomes "decorators" in the system message and conversation content; the original structured_prompt stays unchanged. Template values are not applied inside triple-backtick code blocks. Using template strings in Promptdown allows for more flexible and context-sensitive interactions with language models.

Contributing

Contributions are welcome! Feel free to open an issue or submit a pull request.

License

Promptdown is released under the MIT License.