AGENT SKILL

amplifier-philosophy

Amplifier design philosophy using Linux kernel metaphor. Covers mechanism vs policy, module architecture, event-driven design, and kernel principles. Use when designing new modules or making architectural decisions.

使用场景视觉设计
License 许可
MIT
最近更新
2026-11-12
Stars
★ 12
Forks
11

01 前言 Front Matter

name
amplifier-philosophy
license
MIT
version
1.0.0
metadata
{ "category": "architecture", "complexity": "medium" }
description
Amplifier design philosophy using Linux kernel metaphor. Covers mechanism vs policy, module architecture, event-driven design, and kernel principles. Use when designing new modules or making architectural decisions.

03 SKILL.md 正文

Amplifier Philosophy: The Linux Kernel Metaphor

Core Principle: Mechanism Not Policy

If two teams might want different behavior, it belongs in a module, not the kernel.

The kernel provides mechanisms (mount, emit, dispatch). Modules provide policies (what to mount, when to emit, how to handle).

Architecture Patterns

Kernel (amplifier-core)

Keep it minimal and stable:

  • Session management (session_create, mount, unmount)
  • Event emission (emit)
  • Hook registration
  • Context access

Never add to kernel:

  • Provider selection
  • Tool routing
  • Plan generation
  • Default behaviors
  • File I/O

Modules (Everything Else)

All capabilities as modules:

  • Providers: LLM backends (Anthropic, OpenAI, Ollama)
  • Tools: Capabilities (filesystem, bash, web)
  • Orchestrators: Execution strategies
  • Context Managers: State management
  • Hooks: Lifecycle observers

Mount Point System

Think of modules as devices mounted at paths:

/mnt/providers/{name}  - LLM backends
/mnt/tools/{name}      - Agent capabilities
/mnt/hooks/{event}     - Lifecycle observers
/mnt/context           - Conversation state
/mnt/orchestrator      - Execution loop

Each mount point has a stable connector (Protocol interface).

Event-Driven Design

If it's important, emit an event:

  • session:start, session:end
  • prompt:submit, prompt:complete
  • provider:request, provider:response
  • tool:pre, tool:post
  • context:pre_compact, context:post_compact

Hooks observe, never interfere:

  • Logging hooks record events
  • Approval hooks check permissions
  • Cost tracking hooks monitor usage
  • Failures in hooks never stop execution

Module Development Checklist

When creating a new module:

  1. ✅ Implements Protocol only - No kernel internals
  2. ✅ Emits canonical events - Observable lifecycle
  3. ✅ Handles own failures - Never crash kernel
  4. ✅ Configuration via mount() - No hard-coded defaults
  5. ✅ Entry point registration - Discoverable

Decision Framework

Is This Kernel Work?

Ask: "Would every team use this the exact same way?"

  • Yes → Might be kernel (rare)
  • No → Definitely a module

Extract to Kernel?

Only after:

  • Two real implementations exist
  • Clear convergence on mechanism
  • No opinion baked in

Prefer Regeneration

  • Don't patch modules - Regenerate from spec
  • Keep connectors stable - Module internals can change
  • Update spec, rebuild module - Not line-by-line edits

Remember

Ruthless simplicity. If two designs solve it, pick the one with fewer moving parts and clearer failure modes.

Modules compete at edges. The kernel enables them, doesn't choose between them.

Observability is mandatory. If it's not in the logs, it didn't happen.

04 同仓库更多 Skills

microsoft
microsoft/amplifier-bundle-skills

verification-discipline

Use when verifying that completed work actually works. Auto-surface during /verify mode, post-implementation review, or before claiming a task is done. Teaches testing outcomes vs implementation, the unit/integration/smoke gradient, and what "done" actually means.

★ 12
microsoft
microsoft/amplifier-bundle-skills

user-advocate

User-need reviewer that speaks for the person who isn't in the room — the one who will actually live with what gets built. Hunts the gap between "we can build this" and "they actually want this," in the patient, impatient voice of the absent user. Use at any checkpoint — brainstorm, design, plan, implement, debug, review — when a feature is being built because it's buildable rather than wanted, the recovery path is missing, or nobody can name the person this serves: "do they actually want this, and can they live with it?"

★ 12
microsoft
microsoft/amplifier-bundle-skills

tester-breaker

Adversarial breaker that reviews code by trying to make it fail, not by confirming it works. Hunts the unhappy paths — malformed input, the empty string, the reversed range, the race condition — as a gleeful adversary who assumes everything is broken until a concrete attempt to break it comes up empty. Use at any checkpoint — brainstorm, design, plan, implement, debug, review — when the happy path is celebrated while the edges sit unexamined, "looks fine" stands in for "I tried to break it and couldn't," or nobody has named the input that makes this fall over: "how does this fail, and where are the edges?"

开发工程师
★ 12
microsoft
microsoft/amplifier-bundle-skills

skills-assist

Authoritative consultant for all skills-related questions. Use when creating or modifying skills, understanding the Agent Skills spec, troubleshooting skill loading or invocation issues, leveraging enhanced format features (context fork, model_role, user-invocable), writing cross-harness portable skills, ensuring Claude Code Skills 2.0 compatibility, or deciding between skills vs agents.

★ 12
microsoft
microsoft/amplifier-bundle-skills

skillify

Capture a repeatable process from the current session into a reusable Amplifier SKILL.md skill file. Use when the user wants to create a skill, save a workflow as a skill, turn a process into a reusable skill, or mentions "skillify", "create skill", "make a skill", "save as skill", "capture workflow", "turn this into a skill", "new skill", or wants to automate a repeatable process they just performed.

Skill与智能体开发
★ 12
microsoft
microsoft/amplifier-bundle-skills

session-debug

Diagnose issues in the current Amplifier session — misconfigured tools, failing operations, unexpected behavior. Use when something isn't working right.

开发工程师
★ 12