Microsoft open-sources SkillOpt: training skill docs lifts GPT-5.5 accuracy 23.5 points

sanjaykalra · x · 2026-09-27

Microsoft Research open-sourced SkillOpt, which turns the most valuable asset of an enterprise AI program into a markdown file. Instead of touching model weights, it scores agent runs and iteratively edits the skill document the agent reads, keeping an edit only if it beats a held-out validation set. On GPT-5.5, Microsoft reports a 300–2,000 token trained skill file lifted accuracy by 23.5 points in direct chat and 19.1 points inside Claude Code — though these are Microsoft's own benchmarks, and the author wants them validated on real enterprise workloads. Framed as "Rent the Gym. Guard the Playbook.": the model is rented, but the bestskill.md is yours, and the paper claims it transfers across models and harnesses without retraining. Your validation set becomes the moat.

Original post →

More from coding & agent

coding & agent channel →