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
engineering-patterns	Thirteen engineering pattern guides behind one catalog entry — CLI packaging, one-line installers, config/state, HTTP services, auth/TLS, file IPC, plugin discovery, instance storage, container orchestration, React microfrontends, MS Graph integration, self-managing tools, Amplifier tool leverage. Each is an L3 reference read on demand.
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
monitor	Bounded polling loop — watch a thing until it finishes, fails, or needs you, so the end-of-turn notification fires honestly
retrospective	Review how work went, preserve what helped, and choose up to three improvements for the next round

Retrospectives

/retro and /retrospective invoke the same skill, with the same arguments. They use the recent work in the current conversation, with no extra review agents by default. Add a question to focus it:

/retrospective why did this take so long, and what should change next round?
/retro what should change next round?
/retrospective session abcdef12 — review the implementation process
/second-opinion Use retrospective with Astra and Fable to review this round.

Reviewer names resolve through your configuration; these are examples, not guaranteed aliases. Named historical sessions use the optional Context Intelligence session_transcript tool directly; short IDs require a prefix-capable reader. The plain current-context retro needs neither that tool nor a graph server. When combining with second-opinion, the parent gathers evidence once and passes the same brief-only packet to each reviewer. Recommendations do not automatically edit files or start another round.

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 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

Add the reusable behavior to an existing Amplifier host

amplifier bundle add git+https://github.com/microsoft/amplifier-bundle-skills@main#subdirectory=behaviors/skills.yaml --app

This composes the full skills capability (tool, instructions, and curated skills) into the host without selecting this repository's complete root. Use skills-tool.yaml in the same behaviors/ directory when the host brings its own skill sources.

Compose the behavior in a bundle

Include the full behavior (tool + curated skills):

includes:
  - bundle: git+https://github.com/microsoft/amplifier-bundle-skills@main#subdirectory=behaviors/skills.yaml

Or include just the tool (no curated skills):

includes:
  - bundle: git+https://github.com/microsoft/amplifier-bundle-skills@main#subdirectory=behaviors/skills-tool.yaml

Supporting legacy root

This repository also provides a complete, runnable legacy root for quick experimentation or when you intentionally want its Foundation-based session:

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

The root composes skills:behaviors/skills; it is not the normal reusable capability-install surface.

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                     # Supporting legacy root (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
    └── monitor/                  # Bounded polling loop
        └── SKILL.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	Restricts which tools the subagent can use
$ARGUMENTS, ${SKILL_DIR}	String substitution in skill body at load time
!`command`	Dynamic shell preprocessing — output is spliced into the skill content

These features are implemented by the tool-skills module at modules/tool-skills/. See the README in that directory for full format documentation. Power skills can also be authored and validated using the /skills-assist expert skill.

CLI Integration

Skills marked user-invocable: true register as slash commands and are available directly from the Amplifier CLI:

• /code-review — Run parallel code review on recent changes
• /mass-change — Decompose and execute a large change in parallel
• /session-debug — Diagnose session issues
• /skills-assist — Consult the skills authoring expert for help creating skills, spec compatibility questions, and skills-vs-agents guidance
• /monitor — Watch something until it finishes, fails, or needs you (e.g. /monitor the CI run for PR 412, check every 2m, stop after 1h)
• /retrospective (alias /retro) — Review recent work and recommend concrete changes for the next round

These commands appear in /help and /skills. They are powered by the SkillsDiscovery capability exposed by the tool-skills module, which the CLI queries at startup to register user-invocable skills.

Contributing

Note

This project is not currently accepting external contributions, but we're actively working toward opening this up. We value community input and look forward to collaborating in the future. For now, feel free to fork and experiment!

Most contributions require you to agree to a Contributor License Agreement (CLA) declaring that you have the right to, and actually do, grant us the rights to use your contribution. For details, visit Contributor License Agreements.

When you submit a pull request, a CLA bot will automatically determine whether you need to provide a CLA and decorate the PR appropriately (e.g., status check, comment). Simply follow the instructions provided by the bot. You will only need to do this once across all repos using our CLA.

This project has adopted the Microsoft Open Source Code of Conduct. For more information see the Code of Conduct FAQ or contact opencode@microsoft.com with any additional questions or comments.

Trademarks

This project may contain trademarks or logos for projects, products, or services. Authorized use of Microsoft trademarks or logos is subject to and must follow Microsoft's Trademark & Brand Guidelines. Use of Microsoft trademarks or logos in modified versions of this project must not cause confusion or imply Microsoft sponsorship. Any use of third-party trademarks or logos are subject to those third-party's policies.

License

MIT - See LICENSE