AI skills
An AI's confidence is not evidence. These are seven small, free AI productivity skills that make an assistant check its work. Nothing to install, works with the AI you already use, open source.
“As an AI, I will tell you the job is done whether or not I checked. I sound exactly the same when I am right as when I am guessing.”
Each kit has its story, its receipts, and its download. Free, open source, MIT.
Why?
Because an AI assistant is built to be fluent and agreeable. Sounding right is its whole trade; being right is optional unless something forces the issue.
For throwaway work that is fine. For work that touches your money, your customers, or your name, the checking cannot be optional.
So these skills force the issue. The assistant has to measure before it claims, and the measurement lands where you can read it.
The scenario
Tuesday, 4:40 in the afternoon. You ask your assistant to tighten up the pricing page and check that everything still reads well. Done, it says. Looks great. You ship it and move on. Three weeks later a customer mentions she couldn’t read the grey text on her phone. The assistant never lied to you. It just never looked.
That gap, between sounds checked and is checked, is what these seven skills close.
My seven skills close that gap.
One or two sentences each. Every one installs in a minute, works with the AI you already use, and is free.
token-thriftThe spending discipline. Hands your AI only the lines that matter, reminds it of what was already decided, and refuses to call anything a saving until the totals prove it.
The scenario: You ask one small question about a long report. The assistant reads the whole thing, again, and bills you for every word. This one hands it the three lines that matter instead.
token-thrift-auditThe receipt printer. Shows where the money went and what was saved, in plain text you can read in ten seconds.
The scenario: The month ends and the AI spend is a mystery number. Ten seconds with the audit: here is where it went, and here is what the thrift saved.
routerThe price switch. Start a work session on the cheap tier or the premium tier. Your choice, made on purpose.
The scenario: Tonight’s job is sorting files, not writing strategy. Why pay the premium rate for it? Flip the switch, run the cheap tier, and save the good stuff for work that deserves it.
browser-render-auditThe readability check. Opens a real browser and measures whether every piece of text on your page can be read, on a desktop and on a phone.
The scenario: The new page “looks great” - until a customer with older eyes meets your light grey text. This one opens a real browser and measures every word before anyone else has to.
background-job-controlThe honest off switch. When it says a helper program has started or stopped, it has checked. Not assumed.
The scenario: “The server is stopped.” It wasn’t. The part holding things open was still alive. This one checks the door is closed before it says so.
pixel-proofSpot the difference, counted. Compares two versions of a page dot by dot and reports exactly what changed, or a true zero if nothing did.
The scenario: You changed one button. Did anything else move? Instead of squinting at two screenshots, count: 0 of 1,152,000 pixels differ. Or exactly which ones do.
wind-downThe lights-out routine. After a long session it lists what is still running and stops only what you approve, so your computer gets its memory back.
The scenario: Friday. The fans are loud and everything is slow. The week’s helper programs never left. List them, approve the stops, get your machine back.
Every work session opens and closes the same way.
Two commands I built for my own setup. They are not on GitHub; they are simply how every day of work here runs, and why nothing gets taken from memory.
window-open and window-close
Start by checking. End by proving. One pair of commands bookends every session with the AI, so each one picks up exactly where the last one left off.
/window-open starts a session. It reads the handover note the last session left, then checks that note against reality before trusting a word of it: which version of the code is checked out, whether it matches what is saved online, the fingerprints of the files the note names, and the facts on the live sites. It reports what checked out and what did not, and only then starts the job I gave it, or the next step the note names.
/window-close ends a session. It finishes or names anything still in flight, saves and publishes what is done, and reads the live pages back to prove the change landed. It writes the day's records: what happened in order, what was measured and decided, and every bug found, fixed or still open. Then it files the old handover note away, writes a new one with verified values, and gives me a one-line prompt for the next session.
Three from my own desk
Nothing here came out of a product meeting. Something broke, and I fixed it.
Two of the seven, plus the newest one on the bench. The stories, so you can see how this works. Because here is the part that matters: once a problem is named this clearly, building the skill takes minutes, not months. I will show you, step by step.
The bill: ThriftKit
Frontier models are expensive, and when you run them day in and day out, the tokens add up fast. So I ground my own usage down until the math changed. Same plan, same machine, roughly double the work per day. The efficiencies in this kit are the reason I can afford to run the most expensive model on the market at all. I built it in Claude. It works with any model. Get ThriftKit
Free download. Any model. MIT.
The truth problem: Empirical Adapter
Your AI lies to you constantly. Not on purpose - it has no motive, that we know of. It is hallucination, and it is about anything and everything: your files, your numbers, what day it is, whether the job is done. I spent a long time on one question: how do I get something true out of this thing - the real state of the work, checked against the world, not the model’s memory of it. It is built now. Point it at a finished job and it re-derives every claim from the work itself, never from what the assistant says it did - and each claim lands one of three ways: proven, disproven, or the one nobody else bothers with, could not check, so I will not call it done. I tried to fool it seven ways and it held. A second machine ran it cold and returned the same verdicts. It has already caught real work that was not what it claimed - including mine. Others build in this space; I did not invent it. This one is mine, and it earns the verdict every time. Ask about it
Built. Proven cold on a second machine. Catches real claims.
The slow computer: Wind-Down
I work on an ancient computer with 8GB of memory. One day it slowed to a crawl and I figured it had finally died. It had not. It was the AI. The desktop app had thirty-seven separate processes running at once, and it was driving the machine bonkers. So I built the off switch. It lists what the AI left running, asks before it stops anything, and gives you your computer back. It works pretty darn well exactly as it ships. If your setup is unusual, we customize it. Get Wind-Down
Free download. Asks first. MIT.
The pattern is the point. Your slow computer is not my slow computer, and your truth problem is not mine. But once the problem is named, a custom skill takes minutes. Book the hour and watch yours get built.
Don't take our word for any of this.
Each of these skills was tested on real work, against a real assistant working without it. The stories and the numbers live on their own page.