---

name: microsoft-skillopt description: | Microsoft SkillOpt wrapper for data-driven self-evolving agent skill optimization. Use when the user wants to train, benchmark, evaluate, or systematically improve natural-language skills with validation-gated iterations, best_skill.md artifacts, and trajectory-driven edits using the upstream Microsoft SkillOpt project. author: Microsoft license: MIT category: workflow tags: Microsoft, SkillOpt, Skill优化, 自进化, 评测, 元技能 source: https://github.com/microsoft/SkillOpt

Jimu Platform Adaptation

This skill wraps the upstream Microsoft SkillOpt repository for Jimu users. SkillOpt is not a built-in Jimu tool and should not be invoked as a magic command. Treat it as an optional Python project bundled under upstream/SkillOpt.

• Use Jimu file, shell, todo, skill, and sub-agent tools to inspect skills, prepare datasets, run commands, and review results.
• Do not invent or log API keys. SkillOpt requires explicit model credentials such as Azure OpenAI, OpenAI, Anthropic, or local vLLM configuration; use only credentials the user has configured for the environment.
• Prefer lightweight Jimu-native review first for simple edits. Use SkillOpt when the user needs repeatable train/validation/test splits, validation gates, or empirical optimization loops.
• Keep generated outputs outside source-controlled skill directories unless the user asks to package a specific artifact. Review best_skill.md before replacing any production SKILL.md.
• For Jimu resource-market work, adapt any optimized output back to Jimu frontmatter, registry, i18n/platform wording, and validation rules before committing.

Microsoft SkillOpt

SkillOpt trains reusable natural-language skills for frozen LLM agents through trajectory-driven edits, validation-gated updates, and deployable best_skill.md artifacts. It treats prompt/skill text as a trainable artifact while leaving model weights unchanged.

When to Use

• The user wants to optimize an existing Skill with a measurable benchmark instead of manual prompt editing.
• A Skill has weak activation, inconsistent quality, regressions, or unclear evaluation gates.
• The task requires train/validation/test splits and reproducible skill evolution history.
• The user wants to compare candidate SKILL.md versions and keep the best validated artifact.

When Not to Use

• The task only needs a small wording fix, platform-binding cleanup, or frontmatter normalization.
• No benchmark data, expected outputs, or validation criteria exist yet.
• Required model credentials or Python dependencies are unavailable and the user has not approved setup.

Core Workflow

1. Inspect the target Skill and define what behavior must improve.
2. Prepare a dataset split with train/, val/, and test/ items following the selected benchmark format.
3. Choose the benchmark config and model backend.
4. Run SkillOpt training or eval-only mode from the bundled upstream project.
5. Review outputs/<run_name>/best_skill.md, run Jimu/platform validation, and compare against the original Skill.
6. Apply only the validated improvements needed for the target Jimu Skill.
7. Rebuild and validate the resource registry before release.

Bundled Upstream References

• Upstream README
• Installation
• First Experiment
• Configuration Guide
• Training Loop
• Skill Document Guide
• CLI Reference
• Config Reference
• Rewrite Skill Prompt
• Meta Skill Prompt

Minimal Command Shape

Run commands from upstream/SkillOpt after confirming Python 3.10+, dependencies, dataset paths, and model credentials:

pip install -e .
python scripts/train.py \
  --config configs/searchqa/default.yaml \
  --split_dir /path/to/your/split \
  --azure_openai_endpoint https://your-resource.openai.azure.com/ \
  --optimizer_model <optimizer-model> \
  --target_model <target-model> \
  --out_root outputs/<run-name>

For evaluation without training:

python scripts/eval_only.py \
  --config configs/searchqa/default.yaml \
  --skill outputs/<run-name>/best_skill.md \
  --split all \
  --split_dir /path/to/your/split \
  --azure_openai_endpoint https://your-resource.openai.azure.com/

Jimu Release Checklist

• Confirm the optimized Skill still has valid Jimu frontmatter.
• Confirm references and scripts needed by the Skill are preserved.
• Remove or adapt upstream host-specific tool names before publishing.
• Run the relevant tests or benchmark comparison.
• In the Jimu resource repository, run npm run build && npm run validate before committing.