At 2:30 this morning, a timer fired and a scheduled program picked today's topic for this site, drafted the post, handed it to a second model for a 40-item review, traced the claims to primary sources, and published what you're reading. The first human look usually happens later, in a daily email with the draft attached and a log of every change. That pipeline has published 47 posts since September 9. This post is the whole system in one place: the jobs, the numbers, the failures, and the work we keep.
What AI does here
Nine jobs run the publishing side of this company, and eight of them run on timers with no one touching them while they work:
| Job | When | What it does |
|---|---|---|
| Collect questions | 1:20 AM | Pulls new comments from AI-tutorial YouTube videos into one file |
| Find themes | 2:30 AM | Clusters the comments into topics, ranked by demand |
| Check coverage | 2:30 AM | Compares those themes against posts already on this site |
| Draft | 2:30 AM | One model writes the full post from written instructions |
| Review | 2:30 AM | A second model checks the draft against a 40-item checklist |
| Verify | 2:30 AM | Traces what the draft says about AI tools to that company's own docs |
| Publish | 2:30 AM | Builds the site, deploys it, confirms the page loads |
| Report | 2:30 AM | Emails a human the draft and a log of every change |
| Repurpose | 9:00 AM | Turns the post into an Instagram carousel |
Each job is an agent in the plain sense: a program on a timer that wakes up, does its work, and leaves a log. GLM, an open model, writes the first drafts. GPT, from OpenAI, reviews them. A scraper reads comment sections, and a build command ships the site. This version didn't need a team of engineers; it needed written instructions, a checklist, and the willingness to read the logs. If you copy one job, copy the one that repeats weekly and has a written definition of correct.
Where the topics come from
The scraper visits the comment sections under AI tutorials from five YouTube channels and files each comment with its likes and its parent video. A clustering script then reads the recent pile, drops the replies creators leave on their own videos, and groups the audience questions into themes. Today that was 4,287 comments in 13 themes, ranked by how often people ask, how many likes the questions collect, and how many channels they show up on.
We then check each theme against the posts already published here. An uncovered theme takes the day, and its top question becomes the post. When the current themes are covered, the run falls back to a standing list of topics we wrote, which is where this post came from.
The comments are worth mining because they carry information search trends lack: a question with 26 likes means the topic matters to more than the person who asked it.
What the review catches
The checklist is the quality bar, and it flags the things a first draft drifts into: hype words, predictions stated as certainties, universal claims ("always", "never"), fragments that perform profundity, invented scenarios, and sections that repeat an earlier point in new words. The drafting model then accepts or rejects each flag with a reason, and the log records which. The last few posts took between 9 and 17 accepted fixes per run. A first checklist needs fewer items than ours: ban your hype words, hedge the predictions, and cut anything that repeats an earlier point.
What the human does
We let the models run the repeatable part of this operation, but we keep ownership of the rules. We wrote the brand guide, the voice guide, and the topic list, and we keep reading the drafts after publish. The ban list is the clearest trace: three words got banned in a month, each caught by one of us noticing a pattern in our own drafts, and the checklist grew each time.
Judgment calls happen when the facts move too. In October, a routine check of a company's own help pages showed that a product we were about to explain had a retirement date attached (December 11, verified against OpenAI's migration FAQ that day and again this week). A how-to for that product would have been outdated as soon as it shipped, so the post became a what-to-use-instead piece instead. No checklist catches that; we decide.
What it costs
The compute is the cheap part: a draft, a review, a rewrite, and some web fetches, a handful of model calls per post. In our workflow, attention costs more than compute. Checking a claim against a company's documentation takes longer than drafting the sentence, and a check that skips the source isn't a reliable check. When you price your own version, price the checking hours, not the tokens. What Is an AI-Native Company? How We Built One covers the operating model behind this; the short version: drafting got cheap, so checking is where the quality lives.
Where it breaks
Three failures from the last week, because a system you only see the wins of is a demo:
A review came back truncated (Oct 21). A critique that stops mid-sentence can't be applied. Fix: re-run it, and have long outputs written to a file instead of a chat window.
The coverage flag pointed at the wrong posts (Oct 20). The heuristic that marks a theme "likely covered" matched four posts about business operations, but the theme's top question asked for the costing method itself, which no post answered. Fix: treat every "likely covered" flag as a claim to verify against the post inventory.
The reviewer was confidently wrong (Oct 22). It flagged a product-retirement date as inaccurate. A live check of the company's own help pages showed the date held. Fix, now a rule: when the reviewer and the source disagree, we fetch the source again.
These are ordinary failure modes: outputs get cut, heuristics overreach, and confident models are sometimes wrong. Every step leaves a log, so failures are findable, and findable failures get fixes.
Could you run your work this way?
Three conditions decide whether the pattern transfers:
The job repeats with the same shape. A post a day, an invoice run, a weekly report. One-off creative work doesn't need the machinery.
The output can be checked. By a checklist, a citation, or a test. If you can't state what "correct" looks like in writing, a model can't reliably hit it.
Someone owns the judgment. What to build, what's good enough, what to refuse. Miss this condition and you get confident nonsense on a schedule.
Start with one task that repeats weekly and has a checkable output. Write the instructions, write down the ways it can fail, and keep a log of each run. Add a second task only after the first one survives a busy week without creating more review work than it saves.
FAQ
Does AI write everything on this site?
Each post here starts as a machine draft and gets a machine review, but the rules shaping it (the brand guide, the voice rules, the topic list) came from us, and we read the draft by email each morning. When a post misses, the miss is ours to fix.
What AI tools does GWT Lab use?
GLM drafts, GPT reviews, plus a comment scraper, a clustering script, timers, and an automated site build. The agents' toolset is modest: read files, run scripts, fetch pages, build and deploy.
How much does it cost to run?
The compute for one post is a handful of model calls (a draft, a review, a rewrite, some fetches). Prices vary by model and plan, so the number that matters for your own setup is the shape: checking, not computing, is where the attention goes.
Can a small business run this way?
The three conditions above don't require a big team. Pick one task that repeats weekly, write its instructions down, add a checklist, and keep a log. Grow the number of tasks only when the first one survives a busy week.
What happens when the AI makes a mistake?
It has, three times in the last week (a truncated review, a misleading coverage flag, a wrong reviewer flag). Each one left a rule behind, and the rules live next to the instructions. Mistakes here are expected and logged, which is a different posture from hoping they don't happen.
How do you keep the quality consistent?
A 40-item checklist on every draft, fact-checks against primary sources, and a ban list that grows whenever we catch a new pattern. The checklist catches the standard drift; we catch the new kind.
Want to go further?
If you build your own version, the workflow post is the starting point:
- What Is an AI Workflow? Building Your First One covers the first build.
- What Is an AI-Native Company? How We Built One is the definition behind this system.
- How to Build an AI Content System shows the content pipeline in build order.