Small, sharp agent skills. Free, no signup, no email wall. Each one does a single thing that is genuinely annoying to do by hand.
They also demonstrate how we think a fleet of agents should be run: every skill writes a receipt, and one field on that receipt can only be set by a person.
Paste a Python or TypeScript function. Get back its time and space complexity, the exact line responsible, and the rewrite.
Not "this is O(n²)." Instead: line 7, the .includes() inside the loop, here is the Set version.
It will also tell you when to leave the code alone — if n is the number of weekdays, the quadratic is free and the clear version is better. A recommendation to change nothing is a real finding.
Paste a thread, a transcript, or meeting notes. Get back every commitment you made, who you owe it to, and by when.
The useful part is the ambiguity list: the places you said "I'll take a look at that" — which was thinking out loud to you and a promise to them. It names who is most likely holding you to it, and drafts the one line that would settle it.
It never invents a deadline. "No date stated" is the answer when no date was stated.
games/final-authority/index.html — one file, open it in a browser.
You are the human in the loop for a fleet of AI agents. Six shifts, eighteen proposals, two investigations per shift. Approve, deny, or spend attention looking closer.
You cannot investigate everything. That is the point.
You do not lose by making a bad call. You lose by no longer making calls — approve four in a row without looking and the run ends. The agents never got worse; you stopped being the thing that made them safe.
At the end it shows you your receipt log and the one number that matters: what percentage of your decisions nobody actually reviewed.
Every skill here emits a receipt when it runs — nine fields, append-only, one line of JSON.
{"skill":"hot-path","version":"0.1.0","id":"8f2a91c04e77","run_at":"2026-08-08T19:14:22Z","input":"python, 3 functions, 210 lines","findings":2,"unknowns":1,"recommended":"changed","human":"pending"}input records the shape of the input, never its content. That is the line between telemetry and surveillance.
human starts at pending and only a person ever changes it. Then one query tells you whether your review process is real:
grep -c '"human":"pending"' receipts.jsonlIf that is 96% of your runs, you do not have a human in the loop. You have a rubber stamp with extra steps.
The full convention is in TRACEABILITY.md. Reference implementations in Python and TypeScript, no dependencies, are in tools/receipt/. What exists and where it stands is in REGISTRY.md.
Drop a skill folder into wherever your agent platform reads skills from. They are plain markdown with frontmatter and no runtime dependencies — nothing to install and nothing to trust.
The receipts are not signed. They are an honest record among people who are not trying to deceive each other, and they will not survive an adversary with write access to the file. Provable, hash-chained receipts are a harder problem; our own research into it is unproven and paused. What is here is the part that works today and costs nothing.
hot-path reads structure, not runtime. A profiler measures what happened; this predicts what happens as input grows. If you have profiler output, trust it over this.
Released by classHuman AI LLC for public use. Take them, change them, ship them in your own work. Attribution appreciated, not required.
classHuman AI — a Generative AI Software Engineering, Development & Research LLC. We build and run agentic systems using the best tools the field already has, with humans holding final authority.
LAHA — Love All Humans Always.