SkillForge turns any code repository into an Agent Skill. Paste a GitHub link or pick a local folder, and SkillForge produces a standard SKILL.md that teaches an AI agent how to use that codebase.

How it works: a harness I wrote from scratch runs an agent loop with five tools. A local open-weight model (Gemma 3 4B through Ollama) reads a compact summary of the repo and writes the skill. A deterministic validator then checks the frontmatter, naming rules, length, code blocks, and a grounding check that every function the skill mentions actually exists in the repository. When validation fails, the error message gives the exact fix, so even a small model can repair its own output. The harness, not the model, decides success by re-running validation at the end.

What’s original: the agent loop and step budget, context builder with progressive disclosure, tool registry, JSON retry logic, an API extractor for Python, JS, TS, Go and Rust, the validator, an experience memory that saves lessons from fixed errors and reuses them in later runs, and a web UI with link and folder-upload input and a live log.

Open-weight AI is central: everything runs on your own machine with no paid API. The model writes every skill, and the harness exists to make a small model reliable.

Honest limits: it is slow on CPU, a 4B model can leave placeholders in commands, and coverage of public functions is reported rather than enforced.