My stack
Copy what fits your work. This is the setup I build Agentic Levels with. In October I counted 90 days of my AI work sessions, and most of what I'd installed never ran.
By José Fuentes.Last checked .Next re-test .Usage counts cover Jul 27 to Oct 25, 2026.
What I run every day
Each one holds something the AI can't know on its own: my process, my voice or my checks. Mine aren't public yet, but each card's rule works without them: ask your AI to follow it. Claude Code and Codex, the two AI helpers I use on my files, read the same skill format, and I point both at one folder.
These run before any real work starts. The third, writing-plans, sits under What I borrowed.
Mine. Public link coming soon.
Turns a request into the version a top professional would do, without making it bigger. It asks what the raw material is and how we'll know the result is good.
Why I keep it. Most bad results start with a thin brief. This runs first on every plan, so the plan starts from real material.
Mine. Public link coming soon.
Sends a big plan (three or more stages) to a second AI, OpenAI's Codex, to find holes before any work starts. Two rounds at most, then I decide.
Why I keep it. Two different AIs miss different things. The two-round cap is there because one plan once went 17 rounds before I stopped it.
Mine. Public link coming soon.
The starting point for anything a person will read: page copy, emails, lessons. It picks the voice and writes a draft. Then a separate editor cuts and checks it.
Why I keep it. The writer never grades its own work.
Mine. Public link coming soon.
A list of the words and sentence shapes that make text sound machine-written, and what to write instead.
Why I keep it. Every writer and editor reads this one list first, so the whole setup shares one standard. Helpers read it 329 times.
Mine. Public link coming soon.
The starting point for design work. It finds real references first and shows the page at desktop and phone sizes. Then it puts options side by side and checks them against rules for how people use screens.
Why I keep it. A design judged in your head isn't judged. This looks at the real page every time.
Mine. Public link coming soon.
A reviewer that only judges. It looks at the finished screen and returns ship, revise or hold, held to a 9 out of 10 bar.
Why I keep it. My most-used helper: 386 calls. It never fixes anything, so none of the work it judges is its own.
From Refero
Searches a library of real, shipped app screens and styles before anything gets drawn.
Why I keep it. Starting from real references beats starting from the AI's average idea of a nice page. Its server was called 247 times.
Refero (opens in a new tab)The skill on GitHub (opens in a new tab)
From Paul Bakaus
A design rulebook for type, spacing, color and layout. I added 28 rules for how people use screens and 20 common screen patterns.
Why I keep it. All my design helpers read the same rules, so their feedback agrees.
How every working session with the AI starts, checks itself and ends.
Mine. Public link coming soon.
Wraps up a task. It saves the work and opens the pull request (the request to add a change to the project). Then it writes a plain-English status and updates the hand-off note.
Why I keep it. It replaced the shortcut I typed most: 75 times. Ending every session the same way means nothing gets lost between sessions.
Mine. Public link coming soon.
When a session's working memory is about 60% full, it writes everything the next session needs into one note. Then it opens a fresh session.
Why I keep it. A fresh session with a good note beats a tired one that's losing the start of the conversation. It's also how work moves between Claude and Codex.
Mine. Public link coming soon.
Every question waiting on me goes into one list, with a default answer, in the same turn it comes up.
Why I keep it. Questions buried in chat get lost. One list means I answer them in one sitting. It ran in 31 of my sessions.
Mine. Public link coming soon.
Runs when the same failure happens a third time. It finds the actual cause, proves it, and fixes it so it can't come back.
Why I keep it. Trying the same fix a fourth time is the expensive habit. After three failures, the AI stops and finds out why.
Mine. Public link coming soon.
Sets up a new project folder before anything gets built: the read-me, the hand-off note, and where the project's data lives.
Why I keep it. Every project starts the same way, so any session can find its way around any project.
Mine. Public link coming soon.
Makes a new skill, files it in the right place, and fixes skills that fire at the wrong time.
Why I keep it. When I explain the same thing three times, it becomes a skill. This keeps the new one short and makes sure it fires when it should.
Mine. Public link coming soon.
Reads each message I send in Claude Code and, when it matches well, suggests up to two skills that fit. It only suggests; the AI still decides. It logs what it suggested, never what I wrote.
Why I keep it. I built it because the skill finder I had never fired on its own: 0 uses in 90 days. Replayed over my past messages, it spoke up on about 4 in 100, and about 7 in 10 of its suggestions fit.
It still misses about 3 in 4 of the times a skill would help. On since Nov 7, in Claude Code only; its first live check is Dec 7.
What I borrowed
Big frameworks bundle dozens of skills, agents and hooks. I counted what I actually used from each, kept those pieces as plain files, and dropped the rest.
Writes a plan for any job with more than one step. Each small task gets its own test (a quick check that proves it works). A pre-mortem (imagining the plan failed and asking why) runs before any work starts.
Why I kept it. Of Superpowers' 15 skills, this and systematic-debugging were the two I used: writing-plans ran in 12 sessions. I dropped the plugin itself, which added its intro text to the start of every session, 9,557 times.
Superpowers on GitHub (opens in a new tab)How to use Superpowers
A fixed order for fixing bugs. First, one quick test that fails on exactly this bug. Then guesses about the cause, tried one at a time, then the fix.
Why I kept it. The only debugging method I actually used, on 6 different days. Its first steps now follow Matt Pocock's order.
An independent reviewer at the end of each stage of work. It reads the change and looks for real bugs, risky changes and missing tests.
Why I kept it. oh-my-claudecode ships 19 agents. Its code-reviewer was one I kept calling: 92 calls. So I kept the job as my own reviewer role and dropped the plugin.
oh-my-claudecode on GitHub (opens in a new tab)How to use oh-my-claudecode
Shared words for planning which jobs AI can run on its own: what starts a run, where it runs, and the brakes that stop it. It only runs when I type its name.
Why I kept it. It came out after my 90-day count. I copied it over whole and used it once, to map my own setup. It showed the safest first job to automate was a small daily update I approve, not the big overnight run I'd planned.
Four of his ideas now live inside skills I already run:
Why I kept it. His set had 0 uses in my 90-day count, but the ideas were good. Folding them into my skills added nothing new for the AI to load.
Connectors and plugins I use
I re-checked all three on Nov 14, 2026. Each does a job I'd otherwise do by hand.
From Anthropic
Documents Claude can write and edit inside the Claude app. This site's task list lives in one.
Why I keep it. My most-called connector, mostly from the task list.
From Vercel
Vercel hosts this site. The connector lets Claude check each new version as it goes live, and read the error log when a page breaks.
Why I keep it. Faster than clicking through a dashboard to find out why a page broke.
From Anthropic
Anthropic's own skills, included with the account. They make Word, PowerPoint, Excel and PDF files, guide the AI through a browser or a computer, and help make new skills.
Why I keep it. Anthropic keeps them up to date, so I don't keep my own copies.
Off today: Supabase and Context7 had some use in the 90 days I counted. Neither is connected on my account as of Nov 14. Supabase holds some of this site's data, and Context7 gives the AI current manuals for programming tools.
Worth trying
Try one at a time, on a real task, and keep it only if you use it.
From Vercel
Searches a public catalog of skills when you ask for help with a job, and offers to install one that fits.
Where I stand. In my setup it never fired on its own: 0 uses in 90 days, even with a rule telling the AI to use it. A skill that waits for the AI to remember it may never run. That's why I built the router.
It still helps when you go looking: ask for it by name. A skill from a stranger can run programs on your computer, so ask the AI to read it to you before you install it.
find-skills on GitHub (opens in a new tab)How to use find-skills
Researches what people said about a topic in the last 30 days, on sites like Reddit, Hacker News and YouTube. Then it writes up what it found.
Where I stand. Useful before a meeting, or when a tool changes faster than the AI's training. Posts from the crowd can be wrong, so check anything you plan to repeat.
Reddit and Hacker News work with no setup, and YouTube needs one free extra program. X, TikTok and some others need a key (a password-like code from that service). Some keys cost money past a free allowance.
last30days on GitHub (opens in a new tab)How to use last30days
From Matt Pocock
About 39 free skills for building software with AI. One, grill-me, interviews you about a plan until nothing is left vague. Another, tdd, writes the test before the code.
Where I stand. Most of the set expects a developer's setup, such as an issue tracker (a shared to-do list for software work). So I took ideas instead of installing it. If you work that way, try it whole.
Matt Pocock's skills on GitHub (opens in a new tab)How to use Matt Pocock's skills
What I cut, and why
They're big, general tools, and I wasn't using most of what they loaded. Every skill and plugin adds a short note about itself to every message, used or not. Pile up enough and the tools cut those notes short, so the AI can miss the one you need.
A framework for planning and building projects in stages, with dozens of commands and 12 agents.
Why I cut it. In 90 days I typed one of its commands once, and its agents ran zero times. Its original project has since been archived.
35 skills, 19 agents, 21 commands and a set of hooks that run on every step.
Why I cut it. One of its agents was worth keeping, the reviewer. Its hooks were noise: one that ran before every step was cut off before it finished 153,368 times, about 1,700 a day. The plugin also showed 633 "update available" nags.
A framework that runs work across several AI models: 44 skills, 50 agents and 47 commands.
Why I cut it. I'd already switched it off, yet its reviewer still ran 10 times in 2 sessions. Switched off isn't the same as gone.
From Garry Tan
About 45 skills with their own programs, plus copies for eight other AI tools.
Why I cut it. I'd copied the whole thing into my setup. 1,812 of its lines pointed at fixed folders on my machine, so it broke whenever something moved. I kept links to the few pieces I used and deleted the copy.
Plugins switched on for the whole account: sales, finance, data, design and more.
Why I cut it. Most had zero uses, and several still added sign-in prompts or extra instructions to every session. Now most are off, and the rest are on only in the projects that use them.
The counts are what I had installed in October. These projects grow fast, so check each one's page for today's numbers.
Your turn
You don't need my setup. You need yours, smaller.
I asked two questions about every piece. For skills: would a smart new hire need this written down? Your process, your voice and your checks pass; instructions that teach the AI what it already knows don’t.
For plugins, hooks and connectors: did I use it last month?
Paste this into the AI tool where your skills live, such as Claude Code, Codex or Google’s Antigravity. It switches nothing off until you say yes, and it deletes nothing. In Claude Code, typing /skill-doctor first shows what each skill costs and how often you use it.
List every skill, plugin, hook and connector I have, grouped by where it came from. Give one line on what each does, and how often I've used it lately. If you can't tell, write "unknown". Don't show or copy any keys, passwords or tokens you find, in your reply or in the note. Then ask me about all of them in one numbered list, with your suggested answer for each: - For skills: "Would a smart new hire at my work need this written down?" - For plugins, hooks and connectors: "Did I use this in the last month?" Wait for my answers. Mark every "no" as "off". Show me the off list and wait for my yes. Then switch those off without deleting anything. Save how to switch each one back on in a note file, and tell me its name. Check that none of them still runs. I'll work a week like this and turn back on only what I miss.
Installed nothing yet? You’re ahead. Add one skill when a job keeps repeating, and run this test again every six months: models change, and so does what they already know.
New to these tools? How to use Claude Code