Ell — Prompt Engineering as Code in Python

What it is

Ell is a Python library that treats prompts as versioned, testable, and trackable code artifacts. Instead of managing prompts as strings in config files or databases, you write them as decorated Python functions. Ell automatically tracks changes, enables A/B testing, and provides a visual studio for exploring prompt versions and their outputs.

The library targets ML engineers, prompt engineers, and developers who need systematic prompt management with version history, performance comparison, and reproducibility.

How it saves time or tokens

Prompt iteration without tracking is chaotic: you lose what worked, cannot compare versions, and cannot reproduce results. Ell solves this by versioning every prompt change automatically. The visual studio shows which prompt versions produced better outputs, enabling data-driven optimization instead of guesswork.

How to use

1. Install Ell: pip install ell-ai.
2. Decorate your prompt functions with @ell.simple or @ell.complex.
3. Call the function normally and Ell tracks the invocation, version, and output.

Example

import ell

ell.init(store='./ell_store', autocommit=True)

@ell.simple(model='gpt-4o')
def summarize(text: str) -> str:
    '''You are a concise summarizer. Summarize the following text in 2 sentences.'''
    return text

# Call like a normal function
result = summarize('Long article text here...')
print(result)

# Launch the visual studio to explore versions
# ell-studio --storage ./ell_store

Related on TokRepo

• Prompt Library -- browse and manage reusable prompts
• AI Tools for Coding -- developer tools for AI workflows

Common pitfalls

• Ell stores prompt versions locally by default. For team collaboration, configure a shared storage backend.
• The @ell.simple decorator expects the docstring to be the system prompt and the return value to be the user message. Reversing these causes unexpected behavior.
• Ell's autocommit tracks every function edit. In rapid iteration, this creates many versions. Use manual commits for cleaner version history.