mitkox / SkillOpt

What is SkillOpt?

SkillOpt is a text-space optimizer for LLM agents that trains reusable natural-language skills without touching model weights. Think of it as prompt engineering at scale: instead of manually tweaking system prompts, you run a training loop that executes tasks, analyzes failures, and iteratively edits a skill document until it works better. The output is a best_skill.md file you can deploy directly to any agent. Built in Python with support for local model servers (llama.cpp, vLLM, LM Studio, Ollama) and cloud backends (Azure OpenAI, Anthropic, OpenAI).

Why is it gaining traction?

The key insight is that you can optimize *how* an agent is instructed without retraining the model itself. SkillOpt uses a validation-gated training loop: it proposes edits, tests them against a validation set, and only accepts improvements. This mirrors how you might optimize code through CI/CD. The local-first fork specifically targets developers who want cheap experimentation without burning API credits. The included DotNetDebug smoke test makes it easy to verify your setup end-to-end before running on expensive benchmarks.

Who should use this?

ML engineers building agentic workflows who want to systematize prompt optimization rather than hand-crafting instructions. Researchers benchmarking LLM agents across tasks. Teams with local GPU resources who want to experiment with prompt tuning without cloud API costs. Not for beginners: expect to understand your benchmark format, configure model backends, and interpret training outputs.

Verdict

SkillOpt is a solid concept with a credibility score of 0.85, but the 19 stars signal early-stage software. The local-first fork is well-documented for its specific use case, and the modular benchmark adapters make it extensible. However, most benchmark datasets are not included, test coverage is unclear, and the research paper context is thin. Worth evaluating if you have a specific benchmark and want structured skill optimization; treat it as a research prototype rather than production tooling.