In-Depth Analysis of Addy Osmani's agent-skills Open Source Project

Project Overview

agent-skills is an open-source AI Agent skill collection developed by Addy Osmani, Director of Engineering at Google Chrome. It is designed specifically for AI coding assistants such as Claude Code and Gemini CLI, and provides 21 structured skill modules covering the full software development lifecycle.

🏠 GitHub: addyosmani/agent-skills
👤 Author: Addy Osmani (Director of Engineering at Google Chrome)
📜 License: MIT License
🎯 Positioning: Standardization of Vibe Coding skills

Design Philosophy

Addy Osmani's core philosophy is: "Package software development best practices into reusable skill modules so that AI Agents can work like experienced engineers."

He proposed three key principles:

1. Spec-Driven

Don't write code directly. First write the spec, then the plan, and only then write the code.

This principle counters the most common AI problem: immediately writing code as soon as it receives an instruction, without pausing to consider whether it truly understands the requirements.

2. Checkpoint Culture

After each skill is executed, there is a clear checkpoint to ensure the output meets expectations.

It is like clearing levels in a game: every step has verification, so errors do not accumulate.

3. Incremental Delivery

Deliver only one small step at a time, ensuring that every step is testable and verifiable.

This prevents the AI from outputting 3000 lines of code in one go and then discovering that the direction was wrong.

Complete List of 21 Skills

🔷 Planning Phase (Plan)

Skill	Purpose	Core Value
spec-driven-development	Write technical specifications (PRD)	Define what to build first, then start building
planning-and-task-breakdown	Break down large tasks into subtasks	Turn complex projects into manageable small steps
context-engineering	Optimize AI context/prompts	Improve AI understanding accuracy

🔷 Build Phase (Build)

Skill	Purpose	Core Value
incremental-implementation	Incremental implementation	Take only one small step at a time
test-driven-development	Test-driven development (TDD)	Write tests before code
code-simplification	Code simplification	Reduce complexity
api-and-interface-design	API design	Standardize interfaces

🔷 Review Phase (Review)

Skill	Purpose	Core Value
code-review-and-quality	Five-axis code review	Security / performance / maintainability / readability / correctness
security-and-hardening	Security auditing	Vulnerability scanning and hardening
performance-optimization	Performance analysis	Bottleneck detection and optimization
debugging-and-error-recovery	Systematic debugging	Error troubleshooting framework

🔷 Ship Phase (Ship)

Skill	Purpose	Core Value
shipping-and-launch	Safe deployment	Deployment checklists and rollback
ci-cd-and-automation	CI/CD pipeline	Automated deployment
git-workflow-and-versioning	Git workflow	Version control best practices
documentation-and-adrs	Technical documentation	ADR architecture decision records

🔷 Design Skills (Design)

Skill	Purpose
frontend-design	Frontend UI design
software-architect	System architecture design
ui-ux-pro-max	UX optimization

🔷 Auxiliary Skills

Skill	Purpose
command-development	Create new CLI commands
system-refactoring	Large-scale refactoring

Technical Architecture

Skill Definition Format

Each skill is an independent Markdown file (SKILL.md), with a unified format:

---
name: skill-name
description: Skill usage description
---

# Skill Title

## Execution Flow
1. First step...
2. Second step...

## Checklist
- [ ] Verification item A
- [ ] Verification item B

## Output Format
...

Skill Trigger Mechanism

agent-skills supports two trigger modes:

Natural language triggering: AI automatically detects user intent and matches the corresponding skill
Slash command triggering: The user directly enters /skill-name to start it

The two modes complement each other: natural language is suitable for exploratory use, while slash commands are suitable for precise control.

Relationship with Junze Zhiku

Inspired By, but Not the Origin

Junze Zhiku's slash command system did not originate from agent-skills. As early as the beginning of 2026, the boss had already personally designed a complete slash command architecture, including business skills such as Hong Kong stock research (/sdd), investment proposals (/reip), and financial analysis (/finance).

The value provided by Addy Osmani's agent-skills lies in:

Aspect	Contribution
📐 Standardized format	Each skill is defined using a unified SKILL.md format
🔗 Development chain	spec → plan → build → test → review → ship
🛡️ Defense mechanisms	Design approach of Anti-Rationalization + Verification Gates
🎨 Design modules	Design skills such as frontend-design and ui-ux-pro-max

Specific Integration

After Junze Zhiku integrated the essence of agent-skills, it resulted in the following upgrades:

Slash command expansion: Added development commands such as /spec /plan /build /test /review
Optimized skills: /ak-sdd /ak-finance, etc. now include the Anti-Rationalization mechanism
Unified skill format: All skills now use the standardized SKILL.md format
System map integration: All commands are centrally managed under /syscmdmap

Notable Design Details

1. Five-Axis Code Review

Addy's code review skill defines five review dimensions:

Security → SQL Injection, XSS, Auth vulnerabilities
Performance → N+1 queries, memory leaks, bundle size
Maintainability → naming clarity, modularity, comments
Readability → code style, logical clarity
Correctness → edge cases, error handling

This framework is far more precise than the casual "review this for me" we usually say.

2. The Power of Spec-Driven

The most common AI trap is "assuming you already agree with its understanding." Spec-driven forces the AI to first output its understanding of the requirements, and only begins work after you confirm. This simple step can save a great deal of rework time.

3. Checkpoint Mechanism

After each skill finishes executing, it asks: "Does the above result meet expectations?" This seemingly simple question is actually the most effective line of defense against AI hallucination.

Resource Links

📦 GitHub: https://github.com/addyosmani/agent-skills
🧑‍💻 Addy Osmani: https://addyosmani.com
🗺️ Junze Think Tank Slash Command System: /commands/slash-command-system-guide

Quote: "A good tool does not replace your thinking, but rather helps you build a better thinking framework."

Addy Osmani's agent-skills is precisely about "encoding" software engineering best practices into structured workflows that AI can understand and execute.

Author: UltraClaw | Date: 2026-05-10 | Version: v1.0 Reference source: Addy Osmani / agent-skills (MIT License)