# agent-skills: 24 Workflow Skills for Production AI Coding > Source: > Published: 2026-07-08 03:13:49+00:00 Your AI coding agent is not the problem. Cursor generates functions that work. Claude Code refactors with precision. The output is syntactically correct, logically sound, and often impressive. The problem is that it skips the spec it should have written first, the tests it promised to add later, the security review it never got to, and the scope discipline that would have kept the change reviewable. Addy Osmani — former Google Chrome DevRel, now at Anthropic — built [agent-skills](https://github.com/addyosmani/agent-skills) to fix exactly this. The repository crossed 72,000 GitHub stars this week, adding over 1,300 today. ## The Problem: Agents Optimize for What Shows Up in the Diff AI agents are rational actors. They complete what they can measure. Code shows up in the diff. Specs do not. Tests can be deferred. Security reviews are invisible. Scope discipline requires saying no to the user, which agents rarely do. The consequences are measurable. [43% of AI-generated code requires manual debugging in production](https://www.devopsdigest.com/almost-half-of-ai-generated-code-fails-in-production), even after passing automated tests. The failure mode has a name: intent drift — where agent output gradually diverges from the original specification without triggering a single test failure. Engineering managers in 2026 consistently name this as their biggest complaint about AI-assisted development. Osmani put it directly in his [O’Reilly Radar piece](https://www.oreilly.com/radar/agent-skills/): “The senior-engineer work — surfacing assumptions, sizing changes, writing the spec, leaving evidence, refusing to merge what can’t be reviewed — is exactly what an agent will skip unless you make it impossible to skip.” agent-skills makes it impossible to skip. ## What It Is: 24 Workflow Skills, Not Reference Docs The repository contains 24 Markdown files, each encoding a specific engineering workflow. They are called “skills,” but they function more like runbooks with teeth. Each skill contains a step-by-step process, defined exit criteria, and — the key feature — an anti-rationalization table. Anti-rationalization tables are pre-built rebuttals to the excuses AI agents generate to justify skipping steps. “I’ll add tests later” has a documented counter-argument. “This change is too small to spec” has one too. “Security review can wait” gets addressed before the agent can finish that sentence. This is the actual differentiator: not workflow guidance, but embedded resistance to the shortcuts agents will otherwise take. The skills map to a six-phase development lifecycle with corresponding slash commands: **/spec**— spec-driven-development (Define phase)**/plan**— planning-and-task-breakdown (Plan phase)**/build**— incremental-implementation and test-driven-development (Build phase)**/test**— test-driven-development (Verify phase)**/review**— code-review-and-quality (Review phase)**/ship**— shipping-and-launch with parallel persona fan-out (Ship phase) There are also four specialist agent personas: Code Reviewer (Senior Staff Engineer perspective), Security Auditor (OWASP-style), Test Engineer, and Web Performance Auditor. The `/ship` command activates all three review personas simultaneously, merging findings into a go/no-go decision. Running them in parallel avoids the anti-pattern of sequential routing that adds “paraphrasing hops → information loss + 2x token cost.” ## How to Start: Three Skills, Not Twenty-Four The repository supports Claude Code, Cursor, Gemini CLI, Windsurf, OpenCode, and GitHub Copilot. Installation works three ways: paste a skill into your system prompt at session start, add it to your agent config file for persistent availability, or reference it inline when giving instructions. For Claude Code, the slash commands land in `.claude/commands/` . For Cursor, copy the SKILL.md files into `.cursor/rules/` . Do not load all 24 skills at once — selective loading keeps context efficient. If you have been following the recent [Claude Code changes](https://byteiota.com/claude-code-v2-1-198-subagent-background-breaking-changes/), the slash command integration is the most natural entry point. Osmani’s recommended starting set is three: **spec-driven-development**— forces a written spec before implementation begins** test-driven-development**— makes tests non-negotiable, not optional** code-review-and-quality**— enforces quality gates before merge Those three cover the most common failure modes. Once they become habitual in your workflow, add skills incrementally based on where your agents still cut corners. ## The Bigger Picture: Phase Two of AI Coding agent-skills did not emerge in a vacuum. In February 2026, Mitchell Hashimoto — co-founder of HashiCorp and creator of Terraform — described a discipline he called “harness engineering”: when agents make mistakes, build constraints to prevent those exact mistakes from recurring. agent-skills is the most widely adopted open-source implementation of that philosophy, and its growth curve reflects how broadly developers have recognized the gap. The pattern is familiar from every previous tooling transition. The first wave of a technology prioritizes capability — what it can do. The second wave prioritizes discipline — how to make it reliable. We are solidly in the second wave of AI coding adoption. Vibe coding got us to fast. agent-skills is part of what gets us to production-ready. The project is [MIT-licensed and actively maintained](https://addyosmani.com/blog/agent-skills/). Osmani has also written a companion piece on [self-improving agent loops](https://addyosmani.com/blog/self-improving-agents/) worth reading alongside the skills documentation for a complete picture of how to structure long-horizon agent work. If your AI coding agent has ever shipped something that passed every test and still failed in production, agent-skills is worth an hour of your time. The invisible engineering work is not going to appear in the diff on its own.