Skill-MAS
A meta-skill that splits a task into stages, picks a profile and runs multi-agent work in Codex and Claude Code.
- 59 ★ on the repository
- Last commit
- Apache-2.0
Extended access
Why this level
- The installer changes agent instructions (`AGENTS.md`, Claude Code rules) and the skill triggers delegation to several sub-agents.
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add skills/skill-masDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- The installer changes agent instructions (`AGENTS.md`, Claude Code rules) and the skill triggers delegation to several sub-agents.
This is third-party code. Review the repository files before installing.
What it does
Skill-MAS evolves a single `SKILL.md` file that teaches a model to design and run a multi-agent system in three stages: task analysis, agent design and workflow orchestration. The work is based on an EMNLP 2026 paper. The installed skill picks a suitable profile for the current task and orchestrates the workflow alongside your own skills, while simple questions and small edits stay lightweight. Four profiles are adapted from benchmarks: `drb` for parallel research and report synthesis, `hlemath` for mathematical reasoning with small-case checks, `bcp` for multi-clue search and candidate comparison and `vitabench` for tool workflows with dependencies. If another orchestration skill already runs the task, Skill-MAS adds handoff and verification guidance to it without creating a second team of agents.
Who it is for. People solving research and multi-step tasks in Codex or Claude Code who want the agent to split work and verify the result on its own.
Good fit when
- You need to compare three retrieval approaches and get a justified choice.
- A math problem must be solved with a proof and small-case checks.
- A task calls for parallel research and a synthesized report.
- You need to find a project from a set of clues and verify the sources.
Not a fit when
- You need exact reproducible benchmark results: profiles are adapted to agents.
- Running several sub-agents for a task is too costly.
- Agent instructions must not be changed: the installer edits them.
Example request
Compare three document retrieval approaches, check assumptions against sources and recommend one for our project.Limitations
Python 3.9 or newer is needed for the installer. It writes a managed block into Codex `AGENTS.md` and a `rules/skill-mas.md` file for Claude Code, preserving existing instructions and backing up changed files. A single-project install uses the `--scope project --project <path>` flags. Orchestration with delegation to several agents uses more tokens than ordinary work. The original optimized profiles are adapted to agent tools and differ from the benchmark versions.
How to disable. Remove the Skill-MAS block from `AGENTS.md`, the `rules/skill-mas.md` file and the skill directory, or uninstall the plugin with `/plugin uninstall skill-mas@skill-mas-marketplace`. Exact paths are in `docs/agent-skills.md`.
FAQ
How do I install Skill-MAS?
Clone the repository with `git clone --depth 1` and run `python3 scripts/install_agent_skill.py --agent both`. For one agent use `--agent codex` or `--agent claude`.
Is there a Claude Code plugin variant?
Yes: inside Claude Code run `/plugin marketplace add linhh29/Skill_MAS`, then `/plugin install skill-mas@skill-mas-marketplace`.
Are datasets or keys needed?
No, using it in an agent needs no benchmark datasets, Python packages or separate API keys: your agent tools and access apply.
Which profiles exist?
`drb`, `hlemath`, `bcp` and `vitabench`: research, math, multi-clue search and tool workflows. The skill picks the suitable one itself.
How do I limit installation to one project?
Add `--scope project --project /path/to/your-project` to the installer command.
Discussion
No questions or reviews yet.
Related
SkillsA skill with 25 commands to audit and improve the prompts, tools and agent designs of a project in Cursor, Claude Code and Codex
A catalog of nearly two hundred skills: code delivery, review, debugging, AI governance, product work and agent coordination.
A meta-skill that folds rules scattered across AGENTS.md, CLAUDE.md and Cursor files into one project skill with task routes.
Meta Skill (ms)
Meta Skill
Meta Skill (ms) keeps skills locally in SQLite and Git, searches them with a hybrid index and serves agents over MCP
Team Mode splits a Codex task between an explorer, an executor and a reviewer while the main agent keeps final acceptance
A meta-skill that creates, evaluates, packages and governs reusable agent skills
README in short
The project was released with a paper and shows how a meta-skill evolves: from code and benchmarks to a portable skill for everyday agents. The Codex and Claude Code section appeared on 9 October 2026.