Axe System Prompt

Overview

This repository contains a prompting strategy to enable tool/function calling for LLM workflows that do not support it explicitly.

Currently, it is focused on creating/updating files, but plan to support generic function calling in the future.

How it works

This system prompt instructs the model to render all code using the txtar format. You can then pipe the model's output to something like response2fs.py to turn the txtar output into filesystem operations.

Tests

You can evaluate different models using llm and changing the model used in build_hello_response.sh.

Running tests (M1 Mac)

1. Install llm using uv

    $ uv tool install llm
   

2. Install the MLX llm plugin

   $ llm install llm-mlx
   

3. Download Mistral-Small-24B-Instruct-2501-4bit

    $ llm mlx download-model mlx-community/Mistral-Small-24B-Instruct-2501-4bit
   

4. Run make test

5. You should see Hello, alice! if all went well.