Benchmarks and skill quality | Research Agent Skills

We evaluate whether a skill is clear, selected for the right requests, and improves the resulting work.

Check	What it measures	Command
Static quality	Structure, triggers, workflow detail, examples and references	`uv run tools/skill_quality.py`
Trigger routing	Whether the right skill appears for a task and stays out of near misses	`uv run tools/trigger_bench.py`
Task outcomes	Blind comparison of answers with and without a skill	`benchmarks/task_evals/run.py`
Domain checks	Citation verification and writing revision behavior	`benchmarks/citation_verification/`, `benchmarks/slop/`

See the benchmark methodology and current results. The model based results are pilot measurements; read their sample sizes before citing them.