I Built an AI Content Engine: What I Automate, and What I Refuse To

An AI content engine is a pipeline that carries research through to finished content: idea, script, video. A machine does the repetitive parts and a person holds the judgment. This is the one I actually run for KAIAK, including the stage I tried to automate, watched fail, and did differently.

I'll start with a confession, because it's the honest way in.

For about a month I chased the thing everyone online is selling. Fully automated content. Push a button, walk away, come back to a finished video. And I got close. I built a machine that could take an idea and carry it most of the way to a published clip without me touching it. Then I watched what it made, and I felt a bit ill.

Those breathless threads never show you the important part. They show you the diagram. The boxes, the arrows, the dashboard with its row of green checkmarks. The idea of automation. What they almost never show is the finished video the thing actually produced, start to end, unedited. And on the rare occasion you find one, you understand why they buried it. Grammatically perfect. Structurally hollow. It sounds like every other video the same machine made for a hundred other people that same morning.

So I had to sit with a question I'd skipped past. What was I actually trying to buy?

Not automation for its own sake. That's just a trophy, and a trophy doesn't do anything. I wanted my evenings back. I wanted to stop losing hours to the slog of dragging a decent thought all the way to a finished thing. It's easy to lose sight of that and start admiring the machine instead of the result. Then you've built a very efficient way to produce work you'd never put your name to.

So this is the opposite of a flex. I'm going to walk you through the real engine. Research, idea, script, video. The parts that save me real hours, and the one I nearly got wrong, where I tried to automate the whole video, watched my first attempt fall flat, and learned which piece has to stay mine. One idea runs underneath all of it, and you may as well hear it now. Automate the drag, never the taste.

The context that shaped all of it

I served as a head of school for years. It sounds like a job about children and big ideas, and some days it was. But the hours mostly went somewhere else. Reports, rotas, the emails about the emails. The slow erosion of the time I'd meant to spend on the work that actually mattered. I felt that drag every day for a long time, and I got a little obsessed with a question I suspect you're sitting with too. How much of this could a machine carry, so a person could get back to the part only a person can do?

That question became KAIAK. I help education leaders and people running small businesses do the same thing. Claw back time from admin, and stay a step ahead of what AI is doing to their world instead of getting flattened by it. The promise is short, and it's on everything I make. Less admin. More impact.

My promise for this piece is just as plain. You'll see the actual content engine behind KAIAK, not a demo I staged for a screenshot. At each stage I'll tell you two things. What the machine handles, and where I keep a hand on the wheel. That second column is the whole point. The people getting hurt by AI right now aren't the ones using it. They're the ones who handed over the parts they should have kept.

Four principles, before any tool

I want to give you the four principles underneath all of this before I open a single app. Tools turn over every few months. The thinking doesn't. Get the thinking right and you can rebuild the engine on whatever ships next.

One. Don't automate what you can't do by hand. Almost nobody follows this. If you can't tell a good script from a mediocre one with both in front of you, automating the writing won't save you. It'll just hand you mediocre scripts faster than you can read them. Automation multiplies the judgment you already have. Bring none, and it multiplies nothing. So the honest first question isn't "can a machine do this." It's whether you can do it well enough to catch the machine when it's wrong. If you can't, you're not ready to automate that piece. You're ready to learn it.

Two. Run the whole thing by hand once before you automate any of it. I wrote KAIAK content manually for weeks before I let a machine near it. Did the research myself, clustered the ideas myself, wrote the scripts myself. It felt slow and slightly precious at the time, like turning down a shortcut on principle. But doing it by hand is the only way to find where the work actually lives. I learned the writing was never the slow part. The finding was. You can't automate a process you've never run, because you don't yet know which steps matter and which ones you were doing out of habit.

Three, and this is the one I'd put on the wall. Fix the front end to save the back end. Most people automate the wrong end of the pipe. They automate the output, generate the script, generate the video, then burn hours downstream fixing what came out. Re-prompting, re-rendering, because what went in was vague. Put the effort at the front instead, on the input and on where a human checks it, and the back end mostly stops fighting you. A sharp brief reviewed at the right moment comes out clean. A lazy one turns into a re-revision grind that quietly eats every hour you thought you'd saved.

Four. Automate the laborious, keep the judgment. Every task has two halves. There's the labour: searching, sorting, formatting, moving things from one place to another, the stuff that's mostly volume. And there's the judgment: choosing, tasting, deciding what's worth saying at all. Machines are genuinely good at the first half and genuinely bad at the second. The whole trick is cutting cleanly along that line every time.

If you want one picture to hold all of this, treat the AI as a fast, absurdly well-read assistant with no taste of its own. It'll read four hundred papers overnight and never complain. It'll draft anything you ask. But it's never met your customer, doesn't know your audience, and truly cannot tell whether something is true for you or merely true-sounding. So you never give it the decision. You give it the legwork and you keep the decision.

The pipeline, stage by stage

Here's the shape of it. The line I care about most is the one between the top lane and the bottom lane.

Now let me open each stage.

Research ingest

Everything starts with research. Content grounded in nothing is exactly the slop I was trying to get away from.

It's a small pipeline I run with Claude Code, the command-line version of Claude, on a schedule. It searches arXiv, where the real AI papers land before anyone's written a take about them, and it pulls a set of RSS feeds from the blogs and journals I trust. Four topics I chose on purpose. AI in education, leadership, practical AI, and systems thinking, the overlap between what I actually know something about and what my audience needs. For each thing it finds it writes a short summary and saves it as a plain text file on disk, sorted by topic. No database, just files in folders. It runs every day and piles up. Right now there are around 2,036 of those summaries in there, waiting.

The part that makes people's eyebrows climb is the cost. It's almost nothing. It runs on a Claude subscription I was already paying for. No per-item fee, no API meter ticking in the background. A research habit that would have been a part-time hire runs quietly on money I'd already spent.

What stays mine is the choice of topics. The machine will read anything I point it at with equal enthusiasm, so the entire value of this stage rides on my having picked the right four corners of the world to watch. That isn't a task to hand off. It's the strategic call that makes everything downstream either sharp or forgettable.

The daily brief

Two thousand research items isn't useful. It's a pile, and a pile isn't content. The second stage turns the pile into something I can act on before I've finished a coffee.

Every morning at five, before I'm awake, an email lands with three ideas in it. Three, not thirty. Thirty is just a second pile, and the point is to reduce.

The brief is a small Python script. It ranks yesterday's summaries, clusters the strongest into three ideas, and rewrites each in my voice by handing it to Claude with my voice guide, then emails the three to me through Resend, a plain email service, on a scheduled job.

The part I'm quietly proud of is that none of the three is built from a single headline. Each one is clustered from three to five related items the machine noticed were circling the same underlying point. One headline is a hot take. Five findings pointing the same direction is a position worth twenty minutes of someone's attention. Each idea comes back already in the KAIAK voice, with a hook and a suggested format, so at 5am I'm reading three finished propositions in my own tone instead of squinting at abstracts.

What I keep is the choice of which idea is worth making. The machine hands me three good doors. It can't know which one fits this week. The one that lines up with a conversation I had yesterday, or a shift I'm sensing, or simply the one I've got something real to say about. I open one.

Reply-to-develop

This is my favourite stage. It's where the whole thing stops feeling like software.

I read the three ideas over coffee, decide idea two is the one, and to set the entire rest of the pipeline going I do this. I reply to the email. I type the number two. I hit send. That's the interface. No dashboard, no app, no logging in anywhere. Twenty minutes or so later, a full script for idea two is sitting in my inbox.

Underneath, a second small Python script watches that inbox over IMAP, the standard protocol mail programs use to read your inbox. It runs every twenty minutes, reads my reply, hands the chosen idea to Claude to develop into a complete script, and mails the finished thing back. All the machinery hides behind something I already use every hour of the day. I didn't learn a tool. I answered an email with one character.

And look at what that character actually is. It weighs nothing. One keystroke. But it's the judgment call, the single piece of taste this stage needs. Which idea gets to become real. The machine did every laborious thing around it, and all I contributed was the choice. That's the design working exactly as intended. My whole input is the one thing only I can do.

Script generation

The shape of what comes back matters as much as the words.

The idea gets built into a segmented script on a fixed arc. Hook, promise, payoff, call to action. Every script rides that same spine, partly because it works and partly because a known shape is what lets me automate around it. The machine always knows what it's aiming at. It arrives in short segments rather than one wall of text, each with the on-screen text and a note on how to shoot it. So what lands isn't a slab of prose I then have to figure out how to film. It's most of the way to a shooting script, with the structural labour already done.

The edit stays mine. I read every script and change things. A line that's technically fine but sits half a degree off from how I'd say it. A hook that's competent without being alive. A claim that's true but that I'd want to soften before it went out under my name. The machine gets me a strong draft. It doesn't get me a finished one, and I'd be lying if I said otherwise. That last ten percent, where it stops sounding like a good template and starts sounding like a person, is mine every time.

Video, and the honest bit

This is the stage where I have to be straight with you. Skipping it would make me exactly the person I complained about at the start.

The plan was to close the loop. The finished script goes to an AI avatar tool, HeyGen, a digital presenter reads it, and the whole thing assembles into a video on its own. Script in, video out.

That sharpened the principle more than I expected. When the machine can clear the bar, and here it can, the human part doesn't vanish. It moves up a floor. It stops being "salvage the bad output" and becomes the editorial cut. Which B-roll sits where, what to lean on, what to drop, where the thing needs a breath. The taste got more interesting to do, not less. What's left to me is telling the difference between "this tool isn't ready" and "I used it lazily," because in the output those look identical, and only one of them is the machine's fault.

So there it is. Research, brief, reply-to-develop, script, and a video stage where the machine does the heavy lifting and I own the cut. A person sits on every seam. I'm not apologising for that. I designed it that way.

Why it doesn't come out as slop

The fair question. If I'm using the same models as everyone else, why isn't my output the same slop I opened by complaining about?

It isn't clever prompting. It's the inputs, and the inputs are unavoidably mine. Grounded content can't come out generic, because generic was never what went in.

Start with the research niche, the corner I chose. The two thousand items in my pile aren't the two thousand in anyone else's, because nobody picked exactly those four topics for exactly my reasons. The raw material is already particular before a word gets written. Then there's the voice guide. I've written down, in plain terms, how KAIAK sounds. The words I won't use, the rhythm I want, the promise sitting under all of it. That document isn't decoration. The machine writes against it every time, so the drafts already lean toward my voice instead of the flat, cheerful register every model defaults to when you feed it nothing. And the third input, the one I'm still building toward, is my own notes and the things I've actually heard people say in the room. The closer the input sits to my real experience, the further the output gets from anything a stranger could reproduce.

Here's the point to sit with. Generic in, generic out. The people drowning in slop earned it. They aimed a raw model at "write me a viral video about AI," gave it nothing of themselves, and got back the average of the internet, which is interesting to no one by definition. To make content identical to mine, someone would need to read my exact sources, hold my exact taste, and have lived my exact years in a school. That's not hard to copy. It can't be copied at all, because the input is a person and there's only one of me feeding the thing.

What I refuse to automate

Which leaves a short list. The things I won't automate. Not "haven't got round to." Won't, on principle, even once the tools are good enough, because knowing where your line sits might matter more than anything else here.

Taste and the final call. Every judgment about what's actually good. The machine proposes. I decide. Is this worth saying, is it true for me, is it alive or merely correct. That never leaves my hands. Hand it over and I've stopped making KAIAK and started running the internet's average opinion through my logo.

The last-mile beauty pass. I build everything function-first. Working, useful, structurally sound, honestly a bit plain. Then a human directs the pass that makes it good to look at. Even where the machine can clear the bar, as it now can with video, a person owns the final look. Which cut, which emphasis, where it breathes. Function first, polish last, and the polish directed by hand.

The relationships, and the question of what's worth saying in the first place. The machine can cluster research into ideas, but the reason any of it matters is that real people are on the other end, lending me their time and their trust. You don't automate a relationship. You just quietly damage it if you try. What's worth saying to those people is a human question the whole way down.

Where to start, if you want your own

The whole thing in a breath. Research comes in on its own, across four topics I chose, for the price of a subscription. A brief lands at five with three ideas, each built from real findings and written in my voice. I reply with a number, and a full script comes back twenty minutes later. I edit it, the taste pass, and it heads to video, where the machine does the lifting and I keep the cut.

If you want to build your own, don't build all of it. Please don't. Build one stage. Just the research digest. A single scheduled search across the two or three topics you genuinely care about, summarised into something you'll actually read. Run it by hand first, so you feel where your own time really goes. And keep your hands on the taste from day one, so that when you do automate more, you're multiplying real judgment instead of scaling an empty shell.

That's the honest version. What I automate, and what I won't. I'm building the rest of this in the open, gaps and all. Follow along and I'll show you the next piece when it's real, not when it's hype.

Frequently Asked Questions

What is an AI content engine?

A pipeline that carries research through to finished content: idea, script, video. The machine does the repetitive parts, gathering, sorting, formatting, drafting, and a person holds the judgement about what is worth saying and whether it is any good. The value is in the division, not in the automation.

Should you automate your whole content process?

No. Fully automated content comes out grammatically perfect and structurally hollow, sounding like every other video the same machine made that morning. Automate the labour and keep the taste. The people getting hurt by AI right now aren't the ones using it, they're the ones who handed over the parts they should have kept.

What should you automate first?

The research, not the writing. Running the whole process by hand first is what shows you where the time actually goes, and for most people the writing was never the slow part. The finding was. Build one stage: a scheduled search across the two or three topics you genuinely care about, summarised into something you'll actually read.

Why fix the front of the pipeline rather than the output?

Most people automate the output, then burn hours downstream re-prompting and re-rendering because what went in was vague. Put the effort into the brief and into where a human checks it, and the back end mostly stops fighting you. A sharp brief reviewed at the right moment comes out clean. A lazy one turns into a re-revision grind that eats every hour you thought you'd saved.

What should never be automated?

Anything you can't already do well by hand. Automation multiplies the judgement you bring to it, so if you can't tell a good script from a mediocre one with both in front of you, automating the writing just hands you mediocre scripts faster. The honest first question isn't whether a machine can do this. It's whether you can catch the machine when it's wrong.

Less admin. More impact.