PromptSDK
PromptSDK is a CLI tool for generating client code with appropriate type signature from a folder containing prompts in Markdown format.
Installation
uv add prompt-sdk-python uv run promptsdk --help
Explanation
Prompts are often written in Markdown and can get quite long. Writing them directly into your Python function has some disadvantages. It has no syntax highlighting and also bloats up the Python source file.
def format_prompt(doc_texts: str, doc_types_definitions: str): return """# Input The following documents have been uploaded: {doc_texts} # Your task Classify each page in each document into zero or one of the following document types, based on their content: {doc_types_definitions} """.format(doc_texts=doc_texts, doc_types_definitions=doc_types_definitions) agent.run(format_prompt(...))
A naive solution would be to write the system prompt in an external Markdown file and read it in the Python code.
# Input The following documents have been uploaded: {doc_texts} # Your task Classify each page in each document into zero or one of the following document types, based on their content: {doc_types_definitions}
prompt_path = PROMPTS_DIR / "classifier.md" prompt = prompt_path.read_text() agent.run(prompt.format(...))
The problem is that you need to inspect the Markdown file to know which variables are used in the prompt.
PromptSDK solves this by generating client code from a prompts folder.
ls prompts
# info_extractor.md classifier.md
uv run promptsdk --input prompts --output generated_sdk.py --name SDKfrom .generated_sdk import SDK agent.run(SDK.classifier(...))
Prompt Templates
PromptSDK uses jinja to read the templates.
--- name: "classifier" description: "The text here appears as a docstring in the generated Python code." --- # Example PromptSDK finds variables that are enclosed in two curly braces. The type of the generated function argument is a String. {{ example }}
Configuration
PromptSDK looks at pyproject.toml for configuration options.
When not given it looks at the CLI flags.
# pyproject.toml [tool.prompt-sdk] input_path = "templates" output_path = "generated_sdk.py" use_class = true class_name = "SDK"