Skill Distillation¶
Technique Ryan Lopopolo runs on openai's Codex harness: point the agent at its own session logs and ask it how to use itself better. The output is new skills, new prompt fragments, or refinements to existing workflow Markdown — the agent teaches itself, and by extension the team.
"One neat thing you can do with Codex… just point it at its own session logs to ask it to tell you how you can use the tool better. Introspection — ask it to do things."
This is the mechanism by which harness-engineering compounds over time: every session is a chance to extract a reusable skill, which flows back into the harness and raises the floor for the next task. It is also why Lopopolo advocates 5-10 deep skills over a wide shallow skill surface — depth compounds, breadth fragments.
Pattern fits inside Symphony's "zero layer" of meta-improvement: the agent can cut its own tickets, modify its own workflow MD, and improve the harness that runs it.