SkillOpt
title: SkillOpt type: tool created: 2026-05-28 last_updated: 2026-05-28 related: ["radar/techniques/BuildingSkillsForClaude", "radar/techniques/StructuredPromptDrivenDevelopment", "radar/techniques/AgentsMdInversion", "radar/techniques/LLMEvaluationMethodology"] sources: ["https://microsoft.github.io/SkillOpt/", "https://arxiv.org/abs/2605.23904"] radar_quadrant: Tools radar_ring: Assess radar_position: center
SkillOpt is a Microsoft research artifact that treats a natural-language skill document as the trainable state of a frozen language agent. Rather than fine-tuning model weights or hand-tuning prompts, an optimizer model proposes constrained edits to a single skill.md-style document, gated by held-out validation. Published as an arXiv preprint (cs.AI; cs.CL, v2 dated 2026-05-25) with code at github.com/microsoft/SkillOpt.
Existing approaches to building agent skills — hand authoring, one-shot generation, loose self-revision — do not function as proper optimizers and do not reliably improve from a baseline given feedback. SkillOpt frames the skill as the external state of a frozen agent and applies the rigor typically reserved for weight-space optimization to text-space.
The system runs a training-loop analog over text:
| Stage | Action |
|---|---|
| Rollout | Frozen target model executes tasks with the current skill, recording scored trajectories |
| Reflect | Separate optimizer model analyzes success and failure minibatches separately to identify reusable procedures |
| Edit | Add/delete/replace operations are merged and ranked under a bounded edit budget — a "textual learning rate" |
| Gate | A candidate skill is accepted only if it strictly improves a held-out validation score |
Additional control elements include a rejected-edit buffer (negative-feedback signal), epoch-wise slow updates and a meta skill (longer-horizon optimizer-side memory), and a single deployable artifact (best_skill.md) that adds zero inference-time cost.
The paper claims best or tied-best across all 52 evaluated (target model, benchmark, harness) cells over 6 benchmarks, 7 target models, and 3 harnesses (direct chat, Codex, Claude Code). Headline accuracy gains over no-skill baseline on GPT-5.5: +23.5 (direct chat), +24.8 (Codex), +19.1 (Claude Code). Skills also reportedly transfer across model scales, between Codex and Claude Code, and to a related math benchmark without re-optimization. Baselines compared against include human-written skills, one-shot LLM, Trace2Skill, TextGrad, GEPA, and EvoSkill.
Yifan Yang, Ziyang Gong, Weiquan Huang, Qihao Yang, Ziwei Zhou, Zisu Huang, Yan Li, Xuemei Gao, Qi Dai, Bei Liu, Kai Qiu, Yuqing Yang, Dongdong Chen, Xue Yang, Chong Luo. Microsoft involvement inferred from the aka.ms/skillopt short URL; the paper does not list institutional affiliations on the abstract page.
Placed in Tools / Assess / center.
Center position reflects two countervailing signals. On one side: the methodology is principled — held-out gating, bounded edits, separated optimizer model — and directly addresses a problem most teams currently solve with intuition (see radar/techniques/AgentsMdInversion for a counter-example showing what naive iteration without proper gating produces). On the other side: it is a research artifact with no documented production use, no stated license on the landing page at time of assessment, and headline numbers are author-reported. The Codex and Claude Code harness evaluations make it directly relevant to teams already authoring skills using radar/techniques/BuildingSkillsForClaude, but adoption requires running the training loop, which assumes a labelled benchmark suite and a separate optimizer model — non-trivial setup. Worth tracking and potentially trialling for skills that are already well-evaluated against a stable benchmark; not yet a default replacement for hand-authored skills.
Trial gate: SkillOpt run end-to-end against a personal benchmark suite with a best_skill.md artifact accepted into a live Claude Code or Codex workflow, with measured improvement over the prior hand-authored skill on a held-out task set.
