ruc-datalab/SkillAdam

Give SkillAdam a skill and tell it what you want to improve. It tests your SKILL.md on relevant tasks, learns from the results, and checks proposed changes before updating the file.

Supported platforms: Codex · Claude Code · Cursor Agent · GitHub Copilot.

• Improve skills for your workflow. Describe your goal and provide examples of tasks your skill should handle.
• Learn from earlier attempts. Use feedback from previous revisions to guide the next improvement and address recurring mistakes.
• Make measured changes. Adapt the scope of each revision and keep changes that pass evaluation.
• Stay in your preferred agent. Start an optimization in the agent you already use, with the option to review individual edits.

Demo · Quick Start · Usage · Citation

• [2026.09.08] Our paper, SkillAdam: Stable and Efficient Skill Evolution for Agents, is now available on arXiv.
• [2026.09] SkillAdam is released with integrations for Codex, Claude Code, Cursor Agent, and GitHub Copilot.

• Python 3.10+, with pip and venv, and Git.
• Your chosen agent's CLI, installed, signed in, and available on your PATH: codex, claude, cursor-agent, or copilot.
• For Copilot's default VS Code integration, the code CLI is also required. For Copilot CLI only, pass --skip-vscode-registration to its installer.

Cursor requires the Cursor Agent CLI; the desktop app alone is insufficient.

The default setup uses your existing agent sign-in and needs no local GPU, Docker, or separate model API key. Model calls consume your platform account's quota, and skill and task content is sent to the configured model provider.

Clone this repository:

git clone https://github.com/ruc-datalab/SkillAdam.git
cd SkillAdam

Run one installer for your platform from the repository root:

The installer sets up a dedicated Python environment and registers SkillAdam with your agent. Restart your agent after installation; in VS Code, use Developer: Reload Window.

For more installation options, see the platform guide.

Open the workspace containing your skill and select skilladam-optimize through your agent's skill picker or invocation mechanism. Give it the skill path and your goal:

Use SkillAdam to optimize /absolute/path/to/SKILL.md for writing concise,
actionable code reviews that catch correctness issues and edge cases.

Your agent prepares relevant tasks and scoring rules. SkillAdam tests the current skill, proposes changes, and evaluates the revised version. By default, proposed edits are selected automatically; your skill file is updated only when the revised version passes validation.

You can make your request more specific by describing a recurring problem or including a task example:

This skill often produces long explanations without a clear recommendation.
Focus on making each review comment identify the problem, explain its impact,
and suggest a concrete fix.

SkillAdam currently optimizes one existing SKILL.md at a time. Include any context needed to evaluate your examples: supporting skill resources and workspace files are not automatically available during test runs. See task guidance for details.

To review proposed edits yourself, include this in your request:

Before applying changes, show me the proposed edits and let me choose.

You can accept or reject individual edits. SkillAdam then validates the selected changes before updating your skill.

Ask your agent to resume using the same run directory:

Resume the SkillAdam optimization in /absolute/path/to/the/run-directory.

Keep your skill in version control so you can review its history or undo changes.

For benchmark experiments, follow the reproduction guide: use the formal backend configurations and pass strict preflight before real execution. Generic backend examples are for custom runs. Training and evaluation produce new results; no historical score table is used for comparison.

If you find SkillAdam useful, please cite our paper:

@misc{li2026skilladam,
  title={SkillAdam: Stable and Efficient Skill Evolution for Agents},
  author={Gaoyuan Li and Meihao Fan and Yizhe Liu and Shaolei Zhang and Ju Fan and Siyi Wang and Jiaheng Hou and Xudong Weng and Honghan Tian and Zang Li},
  year={2026},
  eprint={2609.08944},
  archivePrefix={arXiv},
  primaryClass={cs.AI},
  url={https://arxiv.org/abs/2609.08944}
}

SkillAdam is released under the MIT License. Copyright (C) 2026 Tencent. All rights reserved. Third-party attribution is preserved in NOTICE.

Welcome to join the Data+AI Enterprise WeChat group, chat and share ideas with other users.

If you like SkillAdam, give it a GitHub Star ⭐