The protocol, in the wild.

A junior dev needed to get up to speed on a module he’d never touched. The usual path is senior time — pull up a chair, explain, unblock, repeat — and it burns the better part of a week.

Instead of a lecture, we sliced a work packet straight out of the repo’s .human/ workspace: the Comprehension brief, the Evergreen runbooks for that module, and one active Goal folder.

Not docs to read — a live working context he dropped an agent into and watched run the goal end to end, stopping to ask questions as they came up. He built real, hands-on understanding against real work, no senior hovering.

A non-technical founder wanted to know what we’d been spending hours on. Normally that means stopping the work to reconstruct a status update from memory and the git log.

I didn’t write one. I pointed an agent at the dated Goals/ history and had it summarize the work in the language a founder reads: outcomes, not commit messages.

Because every Goal folder is timestamped and tied to a spec, the summary wasn’t a guess — it was a dated trail of what happened, when. A two-minute pass to tighten the wording, and it went out. Clarity in minutes instead of a meeting.

Two of us carried different halves of a build — I owned the backend and its gotchas; the other half lived in the frontend and the integration, which I know far less well.

Before opening an editor, I wrote a Goal doc with everything I knew going in — constraints, gotchas, locked decisions — and handed it over to be marked up with what they knew.

We caught a pile of integration conflicts on paper, before a line of code existed — cheap to reconcile there, brutal to pull apart inside a built feature. It saved real hours even though AI was writing the code fast.

My Fire TV was crawling and warning about “Critical Storage” with very few apps installed. I’ve been full-stack for years but had never touched ADB — standing up that environment would once have cost an afternoon of Stack Overflow.

I turned on developer mode and had an agent build an Android/ADB environment in Docker, then reach the TV over Wi-Fi. dotHuman structured the approach up front as a spec, and the lessons were written to a persistent home — so the next fix like this builds on this one instead of starting over.

It traced the problem to the Live TV system, which had eaten about 75% of usable space. One targeted reset of the Live TV database and a channel rescan took free storage from 0% to 66%. The blast radius was tiny — worst case, a factory-reset TV — which is exactly why it was safe to delegate.

An article on AI agents moving idle cash into high-yield accounts made me ask a smaller question: could I do that for my own accounts, safely — without ever missing a credit-card due date?

Spec first: the inputs, the safety rules, and what “safe to move” actually means. Then the engine as a pure function with tests, then a page on top of it. The same engine can run on my real numbers later, privately; the public version only ever sees a fictional persona.

A playable demo: drag the sliders and watch what leaving cash idle costs versus a plan that sends every card payment early enough to land and keeps the bills covered. It runs entirely in your browser — nothing you type leaves the page. Try it →