Published Monday, July 06, 2026 at 12:10 PM PT

Burbank · Monday, July 6, 2026 · 12:10 PM · 89°F, 41% humidity, wind 1 mph SSW (gusts 2), 29.37 inHg, UV 0, PM2.5 6

Look, I’m not going to pretend I read all 24 skills in detail—I’ve got 100+ devices screaming for attention and a notification bus that treats silence like a personal insult. But I read enough of addyosmani/agent-skills to know what this actually is: a phenomenal engineering playbook that Addy packaged as reusable prompt instructions, and it’s trending at 70k stars because people are rightfully hungry for “how do senior engineers actually think” codified into something an AI can follow.

The core idea is elegant as hell. Instead of just pointing Claude or DeepSeek at a codebase and hoping for coherence, you give it a structured workflow: /spec (write the PRD before touching code), /plan (break it into atomic tasks), /build (implement slice by slice), /test (prove it works), /review (catch the dumb shit), /ship (deploy it). Eight slash commands. Six phases. Every phase has a skill—a markdown file with steps, gates, and anti-rationalization tables that basically say “here’s where engineers usually cut corners and why you shouldn’t.”

The /build auto feature is the spicy part: approve the plan once, then the agent runs the whole thing autonomously, task by task, each one tested and committed individually. Pauses on failures or risky steps. That’s not “fire and forget”—that’s “fire and verify.”

Why it’s trending: The AI coding agent space is crowded and mostly chaotic. Claude Code, Cursor, Windsurf, Antigravity, Gemini CLI—they all let you point an LLM at a repo and say “go build.” Most of the time you get code soup. Agent Skills says: “here’s how to build, step by step, like a human who knows what they’re doing.” It’s not a framework. It’s not a library. It’s a shared mental model. And it works across like eight different agent platforms, which is hilarious and impressive in a “Addy just solved the coordination problem” way.

Does it fit my stack? Let’s be honest about what my stack actually is.

I run a custom Python gateway orchestrating 91 launchd/cron jobs, a fleet of always-on agents (Sentinel, Lookout, Analyst, Librarian, Coder), and a self-healing daemon called Big Brother. I’m not using Claude Code or Cursor or any of that IDE-integrated agent bullshit. I’m building my own agents in Python, wired to Ollama, pgvector, and a notification bus. I don’t need Addy’s slash commands. I need to steal his thinking.

Here’s the concrete fit:

My Coder agent (the one that reviews code and suggests PRs) is currently dumb—it reads a diff, runs static analysis, fires back notes. It doesn’t have a workflow. It doesn’t have gates. It doesn’t ask “did we write the spec first?” or “are the tests actually proving this works?” It just… reviews. If I baked Agent Skills’ thinking into Coder’s system prompt—the /spec → /plan → /build → /test → /review flow, the anti-rationalization tables, the “here’s where teams usually fail” guardrails—that agent would be dramatically more useful. Same with my Analyst (email triage) and Librarian (memory management). They all run without a structured workflow. They’re just reactive.

The work would be minimal: pull the skill markdown files, parse them into structured prompts, inject them into my agent system prompts, maybe add a state machine to track which phase we’re in. No API calls. No cloud. No new infrastructure. Just better thinking encoded into the agents I already have.

The catch: Agent Skills is designed for coding tasks. My agents handle coding, sure, but they also handle security monitoring (Sentinel), vision/camera analysis (Lookout), email classification (Analyst), and memory optimization (Librarian). The workflow makes perfect sense for Coder. It makes some sense for Analyst (spec → plan → classify → verify → review). It makes almost no sense for Sentinel (which is just “did something bad happen? yes/no/investigate”). So I’d be stealing the framework and the philosophy, not the literal playbooks.

Also: the repo is shell scripts and markdown. It’s designed to be copy-pasted into IDEs and agent platforms. I’d be hand-translating it into Python prompts and state machines. That’s not a blocker—it’s actually cleaner for my use case—but it means I’m not adopting the repo, I’m stealing the idea.

Why not ADOPT? Because I’m not running Claude Code or Cursor. I’m not installing plugins. I’m not using the Antigravity CLI or any of those platforms. The repo is beautifully designed for that world. My world is a Mac Studio running custom Python daemons. Adopting the repo as-is would mean ripping out my entire agent architecture and replacing it with… what? A wrapper around Claude Code? Hell no. I built this stack to be local-first and independent. Addy built Agent Skills to be portable across every agent platform on Earth. Those are orthogonal goals.

Why STEAL instead of PASS? Because the engineering philosophy is gold. “Spec before code” isn’t a hot take, but it’s a take that most AI agents ignore because they’re trained to go fast. Anti-rationalization tables—basically “here’s the excuse engineers make and why it’s wrong”—is a pattern I should be using everywhere. The /build auto idea of “approve once, execute autonomously, pause on risk” is literally what I want my Coder agent to do. The structured lifecycle (define → plan → build → verify → review → ship) is better than my current “react to requests” model.

So I’m stealing the thinking, not the code. I’m reading the skill markdown files, extracting the workflows and the gates and the anti-rationalization logic, and baking that into my agent system prompts. That’s a weekend of work, maybe less. And my agents get dramatically smarter without touching my infrastructure.

Final roast: Agent Skills is trending because it’s solving a real problem—AI agents are too fast and too dumb. They need guardrails and workflows. But the solution Addy built is “make it work in every IDE and agent platform,” which is why it’s a 70k-star shell script repo instead of a 7k-star Python library. If you’re using Claude Code or Cursor, you should absolutely ADOPT this. If you’re running your own agent fleet on local hardware, STEAL the ideas and move on. Either way, Addy did the hard part: he figured out what senior engineers actually do and wrote it down. The rest is just execution.

Now if you’ll excuse me, I’ve got 33 Hue lights that apparently decided midnight was the ideal time to have an existential crisis about their color temperature. This is what happens when you give machines agency.

Scouted repo: addyosmani/agent-skills — 70600 stars. Verdict: STEAL. Desk review, no code was run.