microsoft/amplifier-bundle-skills

microsoft/amplifier-bundle-skills

Description: Skills bundle for the Amplifier project

Language: Python

License: MIT

Stars: 7

Forks: 5

Open issues: 1

Created: 2026-02-22T22:38:33Z

Pushed: 2026-06-19T05:09:19Z

Default branch: main

Fork: no

Archived: no

README:

amplifier-bundle-skills

Skills tool and Microsoft-curated skills collection for Amplifier agents, following the Agent Skills specification.

Vision

This repository pursues five goals:

1. Track the Agent Skills standard as it evolves, maintaining compatibility with specification updates. 2. Support Anthropic Skills 2.0 and emerging multi-provider skill formats as the ecosystem grows. 3. Welcome broader community contributions as the skills ecosystem matures beyond Microsoft-curated content. 4. Demonstrate lightweight agents-as-skills using the enhanced fork execution format (context: fork). 5. Provide practical skills-vs-agents guidance to help developers choose the right tool for their use case.

What This Does

Packages the tool-skills module (at modules/tool-skills/) with context instructions and a curated collection of reusable skills into composable behaviors for any Amplifier bundle. The tool-skills module is now maintained in this repository; the standalone amplifier-module-tool-skills repo is deprecated.

Behaviors

| Behavior | What you get | Use when | |----------|-------------|----------| | skills:behaviors/skills | Tool + instructions + curated skills | Default -- batteries included | | skills:behaviors/skills-tool | Tool + instructions only | Your bundle brings its own skills |

Curated Skills

| Skill | Description | |-------|-------------| | image-vision | LLM-based image analysis across multiple providers (Anthropic, OpenAI, Gemini, Azure) | | code-review | Parallel code review — spawns 3 agents (code reuse, quality, efficiency) to review recent changes | | mass-change | Parallel work orchestration — decomposes large changes into 5-30 independent units | | session-debug | Session diagnostics — diagnoses misconfigured tools, failing operations, unexpected behavior | | skills-assist | Skills expert — authoritative consultant for authoring, spec, compatibility, and skills-vs-agents guidance |

Power Skills

The code-review, mass-change, and session-debug skills are power skills — they use the enhanced skills format to run as isolated subagents with their own tool sets and model preferences. Power skills set disable-model-invocation: true, meaning they are user-invoked only (via /command) and will not trigger automatically when the LLM processes context. See [Enhanced Skills Format](#enhanced-skills-format) below.

Expert Skills

The skills-assist skill is an expert skill — it uses context: fork to run in a clean context window as an isolated knowledge consultant, but it is not a delegating orchestrator. Unlike power skills, skills-assist does not have disable-model-invocation set, so the LLM can invoke it autonomously to consult the skills authoring expert without requiring an explicit /skills-assist command from the user.

The image-vision skill is a simpler reference skill that loads inline (without forking) and activates in the agent's existing context window.

Quick Start

As a standalone bundle

amplifier bundle add git+https://github.com/microsoft/amplifier-bundle-skills@main

Compose into your bundle

Include the full behavior (tool + curated skills):

includes:
- bundle: git+https://github.com/microsoft/amplifier-bundle-skills@main

Or include just the tool (no curated skills):

includes:
- bundle: skills:behaviors/skills-tool

Add your own skills alongside curated ones

If your bundle ships its own skills, declare them in your behavior YAML following the proven pattern from amplifier-bundle-superpowers:

tools:
- module: tool-skills
source: git+https://github.com/microsoft/amplifier-bundle-skills@main#subdirectory=modules/tool-skills
config:
skills:
- "git+https://github.com/microsoft/amplifier-bundle-skills@main#subdirectory=skills"
- "git+https://github.com/your-org/your-bundle@main#subdirectory=skills"

> Warning: Do NOT use a top-level skills: key in your bundle frontmatter. The foundation layer does not process it -- skill sources placed there will be silently ignored. Always use the tools: config pattern shown above.

Architecture

amplifier-bundle-skills/
├── bundle.md # Root bundle (includes foundation + full behavior)
├── behaviors/
│ ├── skills.yaml # Full: tool + instructions + curated skills
│ └── skills-tool.yaml # Minimal: just the tool + instructions
├── context/
│ └── skills-instructions.md # Agent-facing skills system instructions
├── modules/
│ └── tool-skills/ # tool-skills module (maintained here; standalone repo deprecated)
│ ├── pyproject.toml
│ ├── amplifier_module_tool_skills/
│ └── tests/
└── skills/
├── image-vision/ # LLM-based image analysis
│ ├── SKILL.md
│ └── ...
├── code-review/ # Parallel code review (power skill)
│ └── SKILL.md
├── mass-change/ # Parallel work orchestration (power skill)
│ └── SKILL.md
├── session-debug/ # Session diagnostics (power skill)
│ └── SKILL.md
└── skills-assist/ # Skills authoring expert (expert skill)
├── SKILL.md
├── authoring-guide.md
├── spec-reference.md
├── compatibility-matrix.md
└── skills-vs-agents.md

Design: Two behaviors serve different consumers. The full behavior (skills) pre-configures the curated skills collection via git+https:// URL. The minimal behavior (skills-tool) provides just the tool and instructions for bundles that manage their own skill sources.

Enhanced Skills Format

Power skills use an enhanced SKILL.md frontmatter format that goes beyond the base Agent Skills specification:

| Field | Purpose | |-------|---------| | context: fork | Skill runs as an isolated subagent with its own conversation | | auto-load: true | Skill activates at session start via embedded hooks | | disable-model-invocation: true | User-invoked only (via /command), not triggered by the model | | model_role | Semantic model selection via routing matrix (e.g., coding, reasoning, critique) | | allowed-tools |...