Autoresearch

Run autoresearch with the var parameter set to the target skill's name. It reads the existing SKILL.md, researches better approaches, generates four variations, scores them on a weighted rubric, and commits the highest-scoring version as a pull request.

Each variation is scored 1-5 on clarity, data quality, output value, robustness, conventions, and improvement. Improvement is weighted 3x, output value 2x, clarity, data quality, and robustness 1.5x each, and conventions 1x, producing a weighted total out of 50.

If all variations score at or below the original on the improvement criterion, autoresearch skips the update entirely and sends a notification stating no improvement was found, so working skills are never downgraded.

Yes. When the sandbox blocks outbound curl requests, autoresearch falls back to WebFetch for URL retrieval. For APIs requiring authentication, it uses a pre-fetch and post-process pattern described in the project documentation.

Autoresearch preserves the skill's core purpose and cannot change tags or var semantics without justification. It also cannot add environment variables not already available in the workflow secrets, limiting changes to what the existing infrastructure supports.