What's this about?
LLM agents are proficient at automated software issue resolution in general, but struggle inside a *specific* repository. SkillForge has agents self-distil project-specific skills to close that gap.
What exactly is project-specific knowledge?
Every codebase has conventions nobody wrote down: which helper to use for dates, why one module must never import another, the test that always fails locally and can be ignored, the deploy step that must run first. A new human engineer absorbs this over weeks by making mistakes and being corrected. An agent arrives with none of it every single time, which is why it can be excellent on a benchmark of isolated problems and mediocre on your actual repository. Self-distillation means the agent derives these rules from its own successful and failed attempts, rather than waiting for a human to document them.
Why is this worth noticing?
- It's the same shape as StateM's runbooks and Chain-of-Experience — turning postmortems into persistent, reusable procedure — applied to the specific case of a codebase.
- It's also precisely what today's top paper says skills are actually good for: procedural anchoring rather than knowledge injection. A project-specific skill is a procedure, not a fact.