harness-builder

Skills for building agentic harnesses: structures where AI agents do the work and humans steer through the harness itself (immutable specs as oracle, generator/evaluator role separation, guard hooks, bounded rounds with human escalation, drift ledgers) rather than through per-iteration chat or PR comments. Covers scoping a harness (one oracle + one evaluator per harness), compiling a source of truth into executable spec artifacts, designing the generate/evaluate loop, writing guard hooks, drift adjudication and escalation, scaffolding a new harness plugin from a skeleton, and maintaining a harness reactively from failures. Domain-agnostic: applies to any realm with an oracle and a mechanized evaluation method (services vs contract tests, pipelines vs snapshot diffs, migrations vs parity checks, UI ports vs recorded journeys). Distilled from one production instance that ported a full iOS app to Android with agents writing ~100% of the code.