mitkox/skillopt

SkillOpt is a tool for automatically improving the instructions you give to an AI agent. When you use an AI agent for a specific task (say, debugging code, answering questions, or filling in spreadsheets), the agent's performance depends heavily on the system prompt or skill document it starts with. SkillOpt treats that document like code to be optimized: it runs the agent against a set of test cases, evaluates how well it performed, identifies where it failed, and rewrites the instructions to address those failures. It repeats this loop across multiple training epochs and keeps the best-performing version. Critically, the model itself is never changed. SkillOpt only edits the text of the instructions. The result is a file called best_skill.md that you can drop into any agent setup that uses the same model. This approach means you do not need access to GPU hardware for training or any ability to modify model weights. The project is designed to work with locally-running AI models through any server that speaks the OpenAI API format (such as Ollama, LM Studio, or llama.cpp). You point it at your local server with two environment variables, pick a benchmark task, and run a training script. Cloud providers like OpenAI, Anthropic, and Azure are also supported if you prefer not to run a local server. The repo includes a simple C# debugging example called DotNetDebug as a self-contained test case so you can verify your setup works without downloading an external benchmark dataset. Several other benchmarks (document question answering, math problems, spreadsheet generation, and embodied navigation tasks) are supported but require their own separate dataset downloads. This is an open-source fork focused on local AI workflows, based on a research paper from 2026. The project is licensed under MIT.