autoresearch-claude-code

Run `npx skills add drivelineresearch/autoresearch-claude-code --all -g -y` in your terminal to install the skill globally.

Anything with a measurable metric: ML model accuracy, code runtime, build size, Lighthouse scores, or prompt eval scores. You just need a script that prints METRIC name=number lines.

It measures a noise floor from repeated baselines, locks the benchmark script so it cannot be edited, and only keeps changes that beat the best result by more than the noise floor.

Yes. Budget caps on runs, seconds, and target metric stop the loop cleanly, and hooks keep it going across context compaction without manual supervision.

No. You state a goal, a benchmark command, and which files may change in plain English, and the agent handles branching, testing, and logging automatically.