What problem does it solve? Over time, an AI research harness accumulates friction: bad default parameters, outdated scaffolding steps, and skill descriptions that fail to trigger. This Skill turns accumulated usage logs into evidence-backed optimization proposals for the harness itself, without ever mutating the skill corpus directly. ## Core Features & Use Cases - Log-Driven Analysis: Reads .aris/meta/events.jsonl to compute frequency, failure, convergence, human-intervention, and model-delta analyses across skill invocations. - Ranked Patch Proposals: Generates minimal unified diffs (including scaffolding deletions justified by model bumps) ranked by expected impact, each citing specific log evidence. - Read-Only Producer with Gated Landing: Stages approved patches to .aris/meta/pending/ for a separate human-invoked /meta-apply skill; cross-model review is advisory only and the binding jury runs at landing. - Use Case: After two weeks of running research workflows, run this Skill to discover that 60% of users override the review threshold to 7/10, then stage a patch raising the default. ## Quick Start Ask the AI to run meta-optimize on all skills to analyze the accumulated event log and propose harness improvements.
