Explore SkillOpt for improving Cursor agent skills from our session data

We use Cursor heavily on our system (and related repos) with AGENTS.md, .cursor/agents/, and shared skills.

Our session transcripts live locally under ~/.cursor/projects/*/agent-transcripts/.

Microsoft SkillOpt looks like a way to improve those skill documents from real usage — with validation so changes only land when they actually help.

Ask: Is this worth pursuing for our setup? SkillOpt doesn’t appear to support Cursor transcripts out of the box, so we’d need some adapter work.

Questions to answer:

• Does SkillOpt fit how we work, or is our existing skill-creator eval loop enough?
• What would we optimize first — AGENTS.md, PR review agents, or domain skills?
• How would we know a skill change is actually better (tests, lint, PR review quality)?
• What’s the privacy/cost tradeoff of sending harvested session data to an optimizer API?

Start by reviewing SkillOpt's documented input and validation flow, then inspect the local AGENTS.md, .cursor/agents/, shared skills, and ~/.cursor/projects/*/agent-transcripts/ layout. Compare those inputs with the existing skill-creator eval loop. Done means a documented fit decision, a scoped Cursor transcript adapter proposal, and concrete evaluation, privacy, and cost criteria.