The 5-Minute Coffee Ritual That Runs My Code While I Walk the Dog
A few weeks ago I clipped a video by Alex Finn titled “Loop engineering is the most important AI concept to learn in 2026.” I expected another generic take on prompting. Instead, it showed me something I’d been doing half-heartedly without a name: turning the build process into an autonomous loop instead of a conversation you babysit. The idea stuck because it matched something I was already living with — my Hermes cron jobs run every morning whether I’m at the desk or not.
Here’s the core shift. Traditional vibe coding works like this: you send a prompt, the AI builds, you test, you send another prompt. You are the bottleneck. Nothing happens when you step away, and you can only move one project forward at a time. Loop engineering inverts that. You become the editor-in-chief, not the typist.
Three Skills, One Loop
The “Finn loop” splits into three skills that run on a timer:
Spec is the only manual step. You give the agent a rough idea, it interviews you with about fifteen questions until it fully understands what you want, then it files a detailed task — with acceptance criteria and non-goals — in an issue tracker.
Build runs every five minutes, grabs the next ready issue from the backlog, and implements it. It lints, typechecks, runs the tests, and opens a pull request on its own.
Review runs in parallel, waits until an issue is “ready for review,” then opens the code in a browser, clicks around, tests it, deploys a preview branch, and pings you on Telegram to come look.
What This Looks Like in My Own Setup
The moment that made it click: the video’s daily rhythm is an evening merge with a 🚀 emoji. Two agents grind all day, and you spend two minutes at night reviewing and merging. I already run multiple Hermes cron jobs every morning that write and publish blog posts, generate images, and scan my vault for ideas — fully unattended. The difference is I’d never wired a build loop the same way, where a task gets picked up, implemented, tested, and shipped without me holding its hand.
For a single-user setup like mine, the tools map neatly: skills replace the three prompts, cron jobs replace the five-minute timer, and Google Tasks works fine as the issue tracker. You don’t need Linear or Vercel to start — you need the discipline to write a clear spec and trust the review step to catch the mess.
The Lesson That Stayed
The most useful idea wasn’t the automation itself — it was the acceptance criteria discipline. A spec that can’t be verified by two different engineers probably isn’t a spec yet. Once I started writing tasks that way, my agents stopped guessing and started shipping. Loop engineering doesn’t remove your judgment; it removes your typing.
Have you tried letting an agent run a build loop unattended, or does the idea of an autonomous process still feel like a leap? I’d genuinely like to know where other people land on this — drop a comment below.
Wat vond je van dit bericht?
