Many teams spend time and compute retraining prompts or skills every time they switch models, tools, or benchmarks. The core problem is that a skill learned in one environment often does not transfer well, forcing repeated optimization cycles and wasted resources. SkillOpt solves this by training a single natural‑language skill document while keeping the target model frozen. An optimizer model reads scored rollouts and proposes bounded add, delete, or replace edits. A held‑out selection split accepts an edit only when the score strictly improves. The final artifact is one file named best_skill.md that can be used anywhere without further optimization.
Experiments show that skills trained in one setting retain a large share of their in‑domain gain when moved elsewhere. For example, a skill optimized inside Codex lifted Claude Code performance on SpreadsheetBench from 22.1 to 81.8, slightly above what Claude Code achieved with its own skill. Across model families, harnesses, and benchmarks, every transferred skill stayed above the no‑skill baseline, with retention ranging from 10 % for reasoning‑heavy tasks to over 100 % for procedural spreadsheet tasks. The retained gain tracks task type: procedural skills such as workbook inspection and formula verification transfer well, while math‑reasoning skills show weaker portability.
The practical takeaway is a one‑time offline training cost that pays off across deployment contexts. The exported best_skill.md file is compact—typically between 380 and 2 000 tokens, built from one to four accepted edits—so it adds no inference‑time overhead. Because the artifact is plain text, domain experts can inspect, version, and audit each change via the accompanying edit reports, offering a transparent alternative to shipping fine‑tuned weights.
Adopt SkillOpt to train a skill once, export best_skill.md, and deploy it wherever your model or toolchain runs, gaining portable performance without repeated optimization.
#AI #ML #Productivity #LLM #Optimization #SkillOpt