How do I install the Skillopt skill?

Run npx skillmds@latest add undermybelt/skillopt in your terminal (requires Node.js), paste this page's agent-chat prompt into Claude, Cursor, or any MCP-connected agent, or download the SKILL.md file and copy it into your agent's skills directory.

Is Skillopt safe to use?

This skill has not completed SkillMD's automated safety review yet. SkillMD never runs a skill's scripts for you; review the SKILL.md before installing.

Which AI agents work with Skillopt?

This skill is tagged as working with Claude Code, Claude.ai, OpenAI Codex. SKILL.md is an open format, so most agents that read a skills directory can load it too.

Is Skillopt free to use?

Yes. Installing skills from SkillMD is free, and the skill stays under its author's original license.

Who published Skillopt?

undermybelt (@undermybelt) published this skill. Their other Agent Skills are listed on their SkillMD profile.

Skillopt by undermybelt — Agent Skill

Name: Skillopt
Author: undermybelt

Skillopt

Use when the user asks about Microsoft SkillOpt, optimizing or training agent skill documents with reflective loops, evaluating skill changes with validation gates, or running SkillOpt experiments. This is a wrapper around the local reviewed source checkout and must not run installs, training, WebUI, or model/API calls unless explicitly requested.

undermybelt c5e69eb 3.1 KB Updated Sep 17, 2026

File contents

SkillOpt

SkillOpt is a Microsoft research framework for optimizing agent skill documents through reflective training loops, validation gates, and benchmark-specific evaluation.

Local source checkout:

~/.hermes/external-repos/SkillOpt
Upstream: https://github.com/microsoft/SkillOpt
Reviewed checkout: 75b5c7f31c040b4e8845877f1f2dd664bf366b11

Safety Contract

Treat this repository as executable research code, not a passive markdown skill pack.

Do not run these without explicit user approval:

pip install -e . or optional extras such as .[alfworld], .[claude], .[qwen], .[webui]
python scripts/train.py
python scripts/eval_only.py
python -m skillopt_webui.app
python -m skillopt_webui.app --share
alfworld-download
commands that source .env, write secrets, or call model APIs

Risk notes from intake:

Requires or consumes Azure OpenAI, OpenAI, Anthropic, or Qwen endpoint credentials for real runs.
Training and evaluation scripts call external model backends and write run artifacts under output directories.
WebUI uses Gradio and can create a public share link when --share is used.
Some benchmark paths execute generated code or subprocesses inside benchmark work directories.

When To Use

Use this wrapper for:

Explaining SkillOpt concepts and workflow.
Inspecting local docs, configs, prompts, and source before a proposed experiment.
Designing a safe SkillOpt experiment plan.
Reviewing generated skill-document changes before promotion.
Mapping SkillOpt ideas onto Hermes skill governance.

Safe Workflow

Read local docs first:
- README.md
- docs/guide/skill-document.md
- docs/guide/training-loop.md
- docs/guide/configuration.md
- docs/reference/cli.md
If a run is requested, ask for explicit approval of:
- backend/provider,
- credential source,
- benchmark/data split,
- output directory,
- whether generated code/subprocess execution is allowed.
Prefer a dry-run plan and config review before any package install or model/API call.
Keep generated artifacts outside source-controlled runtime skill directories unless the user asks to promote them.
Before promoting any optimized skill, require validation evidence from held-out data or a predeclared benchmark gate.

Common Commands

Only run after explicit approval and environment review:

cd ~/.hermes/external-repos/SkillOpt
python scripts/train.py --config configs/searchqa/default.yaml --split_dir /path/to/split
python scripts/eval_only.py --config configs/searchqa/default.yaml --skill outputs/run/best_skill.md --split valid_unseen --split_dir /path/to/split

For read-only inspection, use ordinary file reads and rg; no install is needed.

---name: skilloptdescription: Use when the user asks about Microsoft SkillOpt, optimizing or training agent skill documents with reflective loops, evaluating skill changes with validation gates, or running SkillOpt experiments. This is a wrapper around the local reviewed source checkout and must not run installs, training, WebUI, or model/API calls unless explicitly requested.---# SkillOptSkillOpt is a Microsoft research framework for optimizing agent skill documents through reflective training loops, validation gates, and benchmark-specific evaluation.Local source checkout:- `~/.hermes/external-repos/SkillOpt`- Upstream: `https://github.com/microsoft/SkillOpt`- Reviewed checkout: `75b5c7f31c040b4e8845877f1f2dd664bf366b11`## Safety ContractTreat this repository as executable research code, not a passive markdown skill pack.Do not run these without explicit user approval:- `pip install -e .` or optional extras such as `.[alfworld]`, `.[claude]`, `.[qwen]`, `.[webui]`- `python scripts/train.py`- `python scripts/eval_only.py`- `python -m skillopt_webui.app`- `python -m skillopt_webui.app --share`- `alfworld-download`- commands that source `.env`, write secrets, or call model APIsRisk notes from intake:- Requires or consumes Azure OpenAI, OpenAI, Anthropic, or Qwen endpoint credentials for real runs.- Training and evaluation scripts call external model backends and write run artifacts under output directories.- WebUI uses Gradio and can create a public share link when `--share` is used.- Some benchmark paths execute generated code or subprocesses inside benchmark work directories.## When To UseUse this wrapper for:- Explaining SkillOpt concepts and workflow.- Inspecting local docs, configs, prompts, and source before a proposed experiment.- Designing a safe SkillOpt experiment plan.- Reviewing generated skill-document changes before promotion.- Mapping SkillOpt ideas onto Hermes skill governance.## Safe Workflow1. Read local docs first: - `README.md` - `docs/guide/skill-document.md` - `docs/guide/training-loop.md` - `docs/guide/configuration.md` - `docs/reference/cli.md`2. If a run is requested, ask for explicit approval of: - backend/provider, - credential source, - benchmark/data split, - output directory, - whether generated code/subprocess execution is allowed.3. Prefer a dry-run plan and config review before any package install or model/API call.4. Keep generated artifacts outside source-controlled runtime skill directories unless the user asks to promote them.5. Before promoting any optimized skill, require validation evidence from held-out data or a predeclared benchmark gate.## Common CommandsOnly run after explicit approval and environment review:```bashcd ~/.hermes/external-repos/SkillOptpython scripts/train.py --config configs/searchqa/default.yaml --split_dir /path/to/splitpython scripts/eval_only.py --config configs/searchqa/default.yaml --skill outputs/run/best_skill.md --split valid_unseen --split_dir /path/to/split```For read-only inspection, use ordinary file reads and `rg`; no install is needed.

Frequently asked questions