autoresearch-skill

Skill optimization through measurable experiments requires establishing a baseline and running a binary-evaluated, one-mutation-at-a-time loop to ensure every change measurably improves trigger accuracy and workflow clarity.

A binary evaluation score converts success criteria into stable yes/no tests, eliminating subjective bias when refining prompt instructions until the pass rate for specific formatting rules reaches 100%.

To stop regressions when applying mutations to a complex agentic task, isolate the impact of specific changes using a one-mutation-at-a-time experiment loop that keeps only score-improving mutations with non-regression guard checks.

Iterative benchmarking on existing skills requires local read, edit, and shell access to manage experiment artifacts, render dashboards, and execute non-regression guard checks throughout the optimization loop.

Bun is the required dependency to manage experiment artifacts for skill refactoring, supporting the scripts and assets needed to render dashboards and execute the baseline-first evaluation loop.

The best way to improve trigger accuracy in an existing AI agent workflow is applying a structured, baseline-first experiment loop that isolates specific prompt mutations and validates them against binary success criteria.