What Agent Skills actually does
Agent Skills is a collection of markdown documents that get loaded into an AI agent's context window. Each skill is a concise guide on a specific engineering task: writing tests, doing code review, debugging flaky tests, handling errors, refactoring safely, and 30+ others. When you point your AI agent (Claude Code, Cursor, or similar) at the skills directory, the agent follows the documented practices. The genius is that the skills are written for AI agents to read, not for humans. The instructions are explicit, step-by-step, and include examples. The repo was created by Addy Osmani (Google Chrome engineering lead) in February 2026 and hit 77K stars by July. The growth rate (1-2K stars per day) is unusually high for a documentation-only project.
The skills that actually changed my workflow
I tested 8 skills over 4 months. The 3 that changed my workflow: (1) Test-Driven Development: AI agent writes failing tests first, then code, then refactors. My test coverage went from 40% to 78% on saas.pet. (2) Code Review: AI agent reviews its own diff before committing. Catches 30% of bugs I would have caught in code review 5 minutes later. (3) Incremental Changes: AI agent makes small focused commits instead of one giant 500-line commit. Easier to review, easier to revert. The other 5 skills (error handling, refactoring, debugging) are useful but not life-changing. The 3 above are why I keep Agent Skills in my permanent workflow.
How to integrate with Claude Code and Cursor
For Claude Code: add the Agent Skills repository as a submodule, then reference the skills directory in your CLAUDE.md. The agent automatically loads relevant skills based on the task. For Cursor: the same approach works with .cursorrules pointing to the skills directory. For other agents (Cody, Windsurf, Continue): the skills are plain markdown, so any agent that can read files can use them. Setup time: 10 minutes. The skills are MIT-licensed and you can fork and customize them for your team. I added a saas.pet-specific skill that tells the agent about our data pipeline and deployment process.
The 5 skills that disappointed me
Not all skills are equally useful. The 5 that disappointed me: (1) Advanced Refactoring: too theoretical, my AI agent produces the same code either way. (2) Microservices Architecture: the skill is too high-level, my agent does not have the context to apply it correctly. (3) Kubernetes Deployment: skill is outdated, references kubectl commands that have changed. (4) Database Migration: skill is too generic, my project uses SQLite which is not covered. (5) Documentation Generation: skill produces verbose comments that do not match my project style. The lesson: skills work best when they are specific to your task. Generic 'best practices' skills are too vague to be useful.
Who should use Agent Skills
Use it if: you are using Claude Code, Cursor, or similar AI coding agents daily, you find that your AI produces code that 'works but is not great' (no tests, no error handling, no documentation), and you want to standardize how your AI agent works. Skip it if: you are using AI for one-off questions (not as a coding partner), you are happy with 'works on my machine' code, or you have a small codebase where the AI has full context anyway. For most developers doing serious AI-assisted development, Agent Skills is the missing layer between 'AI writes code' and 'AI writes production code.' The 77K stars and Google engineering team endorsement are strong signals that this works.
