upskill — Distill Agent Failures Into Validated Skills

Keyword: upskill · flash to pro · teacher student distillation · ralph loop skill validation

HKUDS/UpSkill (MIT) turns a cheap/weak "Flash" model into a "Pro" performer without a model upgrade. When a session fails, it captures the full context, has a strong Teacher model analyze the failure and draft a skill, then validates that skill against the weak Student model in a closed Ralph Loop (up to 3 rounds) before storing it. Validated skills auto-inject into every future session via a CLAUDE.md index (always in context) plus a full SKILL.md loaded on demand. On Terminal-Bench 2.0, a Flash model + UpSkill (51.6% pass rate, $0.04/task) beat the Pro model it was validated against (50.0%, $0.06/task) — a 41% lower cost result documented in tb_harbor_2.0/RESULTS.md.

This skill is the routing-first wrapper: it installs the tool, wires the three roles (Daily / Teacher / Student), and walks the daily capture → build → validate → serve loop.