Google engineer Addy Osmani releases "Agent Skills": AI programming agents require engineer discipline
Google engineer Addy Osmani released the open source project "Agent Skills", which provides 21 standardized skills for AI programming agents, covering the entire process of specification writing, test-driven development, and code review. GitHub has received 27,862 Stars.
Core conclusion
On May 3, 2026, Google engineer Addy Osmani released the open source project "Agent Skills", which provides a complete engineering skills framework for AI programming agents. The project is buzzing on Hacker News and has received 27,862 stars on GitHub.
Core claim: AI programming agents are junior engineers who are extremely capable but have no engineering intuition - they are good at writing code, but naturally skip the hidden key work of senior engineers: writing specifications, designing test cases, doing code reviews, and controlling the scope of changes. The Agent Skills project attempts to put the “scaffolding of senior engineers” back into the decision-making process of AI agents.
Key Points
- Release time: 2026-05-03
- Affected objects: AI programming tool users (Claude Code / Cursor / Copilot / Codex), automation developers, technical team leaders
- Core changes: From "making AI write fast" to "making AI write well" - constraining AI agent behavior through a standardized skills framework
Background and trigger events
As AI programming tools such as Claude Code, Cursor, GitHub Copilot, and OpenAI Codex enter the mainstream, a common dilemma has arisen in the developer community: AI-generated code is extremely fast but of varying quality. The code can run, but it lacks test coverage, security review, and architectural thinking.
Addy Osmani, an engineering manager on the Google Chrome team, wrote in a blog post: "AI programming agents are junior engineers who are extremely capable but have no engineering intuition - they have no idea of work beyond code diff."
This sentence received 82 discussion points on Hacker News, and a large number of developers shared similar experiences: AI-generated code performs well in the rapid prototyping stage, but requires a lot of manual repair in the production environment.
SEO: AI Programming Agent, Agent Skills, Addy Osmani, AI Code Quality GEO: starting with TL;DR, precise number + date extra points
Agent Skills six life cycles
The project has designed 21 skills around the complete development process of the AI agent, distributed in six stages:
| Stage | Core Skills | Purpose |
|---|---|---|
| Specification | spec-driven-development, api-and-interface-design | Let AI think about what to do before writing code |
| Planning and Design | incremental-implementation, scope-discipline | Break down tasks and control the scope of changes |
| Build implementation | test-driven-development, code-simplification | Write tests first and then write implementation |
| Test Verification | code-review-and-quality, security-and-hardening | Automated code review and security scanning |
| Release Delivery | ci-cd-and-automation, changelog-generation | Continuous integration/deployment automation |
| Continuous Optimization | performance-optimization, technical-debt-management | Performance optimization and technical debt management |
Each skill provides standardized slash commands, including /spec, /plan, /build, /test, /review, /ship, /code-simplify, etc., which can be installed directly into Claude Code or copied to any AI programming tool as a Markdown specification guide.
Five design principles
Osmani set five core design principles for Agent Skills:
- Process is better than document - Workflow should have checkpoints, not long talks
- Anti-rationalization form - Common AI excuses have pre-written rebuttals (such as "This task is too simple and does not require specifications" → "The acceptance criteria still apply")
- Verification is non-negotiable – Every skill must ultimately have verifiable evidence
- Progressive Loading - Load corresponding skills according to the development stage, not all at once
- Scope Discipline - "Only move what you are asked to move."
Among them, the "anti-rationalization form" has attracted the most attention - Osmani calls it "pre-written lies that the AI agent has not yet spoken." For example, when an agent tries to skip code review, the form will tell it directly: "Code review is not optional, and even simple changes require a second pair of eyes."
Three ways to use
Agent Skills supports three access modes:
| Mode | Applicable Scenario | Operation Method |
|---|---|---|
| Direct Install | Claude Code users | Install from GitHub marketplace |
| Markdown Import | Cursor/Gemini CLI/Codex users | Import skill files as spec guides |
| Pure reading | Technical team management | No installation, read directly as team best practices |
This flexible design makes the project not only a tool, but also a methodology.
Implications for AI automation workers
The value of Agent Skills extends beyond Claude Code users—it reflects an important trend shift in the field of AI programming agents: from the pursuit of speed to the pursuit of quality.
For developers using n8n, OpenClaw, or Make.com to build automated workflows, the same principles apply:
- Automated processes should also have specifications and testing links
- AI-generated workflow configurations require manual review
- Scope control reduces unexpected API calls and fees
Related extended information
Tool entry
Tools and platforms covered in this article: Claude Code, ChatGPT, GitHub, Cursor, n8n, OpenClaw, Hermes Agent
Internal link guidance
Topic hub
AI Coding Tools Hub (2026)
From Copilot pricing changes to Claude Code + DeepSeek cost-saving setups—one place to compare tools, read explainers, and follow tutorials.
Explore AI Coding Tools Hub (2026) →Monetization angle
How can you make money from this trend?
WayToClawEarn focuses on verified earn playbooks—not just news. Start from these cases.
DeepSeek + Claude Code Micro SaaS
Run multiple small products on cheap inference
Claude Code bug bounty
Productize agent skills into security services