dojolisted

# Dojo — Skills Earn Their Place ## Overview A skill is a claim about future agent behavior. Untested claims ship as liabilities: self-generated, unverified skills measurably make agents *worse* (SkillsBench −1.3pp), while the same authorship under scored rollouts and validation gates lifts frozen-model accuracy by ~20pp (SkillOpt). It was never agents-writing-skills that failed — it was writing without verification. The dojo is where the verification happens: no skill ships on vibes. Dojo covers the full lifecycle: create, edit, test, evaluate, package. For small edits, run only the kata that the change touches (a description tweak needs kata 6–7, not a full baseline) — say which kata you're skipping and why. ## The Measurability Rule Automated optimization needs a scalar grader; most skill output quality has none. So dojo never pretend-measures quality. Pass criteria are: - **Observable process checks** — y/n facts about what the agent did ("scoped search to a discovered venue before deep-reading: y/n") - **Exact-match trigger tests** — which skill did the router pick Holistic output quality stays human/agent judgment, applied openly as judgment. If a skill someday has a real scalar grader, point SkillOpt at it instead of the dojo. ## Tiers — Match Rigor to Skill Type Full TDD on a reference doc is theater. Classify at intake: | Tier | What it is | Example | Rigor | |---|---|---|---| | **Discipline** | Rules agents rationalize around | verification gates, TDD |