This is a submission for the MLH x DEV Writing Challenge
What I Built
I run several small websites on my own. No team, no sprint board, no one to hand a ticket to. Just me, an AI workbench, and a list that kept growing with the same mistakes happening twice.
At first my workflow was me explaining myself. Before every page I would say: one h1 only, do not lazy-load the hero image, generate the sitemap from the data source, and above all, do not publish anything until I say the word. It worked, until the day it didn't. I asked for an inner page to be written, and it was read as "publish it now" — the page went live before I had confirmed a thing. Nothing broke, but something more useful happened: I stopped explaining and started writing the rules down.
WorkBuddy, Tencent's AI office workbench, lets me save those rules as skills — plain files that get read before any work starts. So I wrote nineteen of them. One governs how a site gets built. One governs how inner pages are written. One governs the launch checklist. One governs what happens when a submission hits a verification wall: stop, hand it to a human, never try to solve it. Every one of those skills exists because I once lost an afternoon to exactly that mistake.
That last one is the whole idea. I do not automate judgment. I automate repetition.
The daily rhythm is quiet and stubborn. At 10:30 every morning a scheduled task picks one asset per project, submits it, and records what actually happened. A weekly inspection reads the traffic data and writes a report. Another line publishes content at 19:00. When something needs a decision — a paywall, a verification wall, a link that might be toxic — the machine stops and writes me a card instead of guessing. I even keep a rule that a script is never allowed to declare a page "passed"; it can only fetch the evidence, and I read the evidence myself.
There was a day I published an article and discovered the URL I had pasted as plain text had never become a link. The page existed. The backlink did not. From then on the check became "is there an actual anchor tag pointing at my site in the DOM," not "did the post go up." That single distinction is now written into the workflow.
Somewhere along the way I realised the notes were more useful than the sites. So I put them on WorkBuddy Central — an independent, unofficial guide that records the parts the official pages leave out, marks which parts I verified and which I could not, and dates every correction instead of quietly overwriting it. It is not a product launch. It is the by-product of working this way for a few weeks.
Demo
There is no demo video, so here is the walkthrough instead.
A mistake happens — say, a link that renders as plain text. I write one sentence about what the correct check should be. That sentence becomes a rule inside a skill. The next morning the scheduled task reads it before touching any site. At the end of the day the ledger records what was submitted, what went live, and what got retired. Every Friday the inspection reads the numbers. Once in a while I take the accumulated notes and publish them on WorkBuddy Central.
Nineteen skills, seven sites in active rotation, fifteen dated logs, and one website that exists because I got tired of re-explaining myself. The whole thing runs on a schedule; the only part that is not automated is the part where I decide.
Partner Technologies
I want to be straight about this: I did not use any technology, credits, or sponsorship provided through an MLH partner program, and I am not going to dress anything up as if I did.
What I did use is WorkBuddy itself, including its project-sharing feature — several people can work on the same project online at once. That matters more than it sounds. The skills and the ledgers are not private notes on my laptop; they live in a shared project, so the rules survive me having a bad day and anyone joining can pick up the same playbook. It is a collaboration feature of the tool, not an MLH partnership, and I am naming it as what it is.
The rest is deliberately unglamorous: scheduled tasks for the daily grind, a static site generator for the pages, Cloudflare for hosting and DNS, Google Search Console for checking whether a page was actually indexed, and a handful of spreadsheets that serve as ledgers. The interesting part is not any of these tools. It is the boundary I wrote around them: the machine may submit, fetch, and record; it may not decide, verify, or publish. Every skill that encodes that boundary came from a specific failure, and I can tell you which afternoon each one cost me.
Hackathon Experience
I went to a hackathon, and the thing I kept afterward was not the demo or the deadline. It was the habit of writing things down while they were still fresh — the bug, the workaround, the reason we chose one approach over another — because everyone knew that by Monday nobody would remember.
I took that habit home and made it mechanical. My projects do not end when a weekend ends; they show up again at 10:30 the next morning, and the next, with a log entry that says exactly what was tried and what happened. The skills are that same instinct, written in a form the machine has to read before it starts. Nineteen of them later, the most useful thing I built during that stretch was not a site. It was a way of not having to learn the same lesson twice.
What I remember most is the moment a rule I had written down caught a mistake before I made it again. No applause, no podium — just a task that stopped and asked me instead of guessing. That is the part that stuck.
Top comments (0)