Smol Developer: Generating a Small Python Project From a Prompt

Smol developer is a project generator that I think of as a junior developer: give it a specification and it writes files for an application. I show a small project and then walk through the setup. This is a May 2023 demonstration; repository commands, dependencies and API interfaces may have changed since then.

My prompt asks for a Python program that reads an API key from an environment file, prompts the user, sends a request to the ChatGPT 3.5 Turbo API and displays the response before asking again. I also request a requirements.txt file and provide an example API call. The tool creates the project structure rather than merely returning a code snippet.

I run the generated client and ask for a joke and a historical fact. It returns responses and accepts another prompt. That shows the example worked in this limited run, not that all generated code is correct, secure or ready for production.

The GitHub project is smol-ai/developer. I compare a detailed prompt to a specification handed to a junior developer: the more clearly the intended files and behavior are described, the more context the generator has. I clone the repository, open its folder in VS Code and configure an environment file for an API key.

The version shown uses Modal as one way to run the generator. I install the Modal client and authenticate when prompted; the repository also describes installing requirements and running main_no_modal.py without Modal. These are historical instructions, so check the current project documentation and provider terms before attempting to reproduce them.

For a simple first test, I run the generator with a prompt asking for a Ruby script that counts to 100. It reports a file to create, uses Modal containers and writes the result into the generated directory. I inspect the produced file and find that it matches the simple task.

For longer specifications, the displayed README allows a Markdown prompt file instead of placing the entire description in the terminal. The README also says an ordinary run deletes and rewrites the generated directory, except for images, while a file-specific option can target a file. Preserve any work you need before rerunning and verify the behavior of the version you use.

I show an example detailed specification for a browser extension that summarizes pages using Claude. The point is that the prompt can describe files, interface elements and behavior in more detail, much like a handoff to another developer. The demonstration does not establish that the extension works without testing.

For the Python example, I inspect the generated environment file, main script and requirements file, configure a key and run the program. It answers a joke prompt and waits for the next request. Keep credentials private and out of shared repositories, and review generated code and dependencies before running or distributing it.