I Made an AI Write the Manual for a Time Machine. Twice.

The second draft was better. The reason why is the thing nobody mentions when they tell you to "just use AI."

You've done this.

You opened ChatGPT, or whichever one your nephew installed, and asked it to write something for the business. The job posting. The thirty-days-late invoice email. The FAQ.

And it was fine. It was fine. It read like a polite stranger who'd heard of your industry once, at a party.

So you decided one of two things: AI is overhyped, or you're bad at it.

Both wrong. I have a time machine to prove it.

Here's the whole post in two sentences, so you can leave now if you want. The AI isn't generic because it's dumb. It's generic because it is a map of the entire world, and nobody has drawn your business on it.

You could. That map is in your head. It's the thing you'd sketch on a whiteboard for a new hire, and you've never written it down.

What follows is what happened when I drew one and handed it over. You can do the same thing this afternoon. It's free.

The experiment

I needed a subject the AI already had strong opinions about. Not opinions the way you have opinions. Opinions the way the internet has opinions: thousands of pages of fan wikis, forum arguments, and a Wikipedia article with more citations than most medical journals, all about something that does not exist.

I picked the flux capacitor. From Back to the Future. The thing in the DeLorean that lights up and makes time travel possible, according to a man who hit his head hanging a clock.

The request was one sentence: write a technical document for Doc Brown's flux capacitor.

I sent that request twice.

The first time, that's all it got. Just the sentence.

The second time, I also pasted in a map: a few hundred rules on how to write instructions a human can actually follow, pulled from Google's technical writing course and a couple of other well-regarded guides, along with how those rules relate to each other.

Same request. Same AI. The only difference was whether it had the map.

The technical term for the map is a knowledge graph, and before I show you what it did, I owe you a plain-English explanation of what that is and why it works. It's the part that matters and the part everyone skips.

Okay, what's a knowledge graph

Here's the thing nobody tells you about how these AIs work, and it's the whole reason this post exists.

The AI does not contain facts. It contains relationships. Trillions of them. It read the internet, and what it kept was not the sentences. It kept the web of "this goes with that," weighted by how often it saw the two together. "Flux" leans hard toward "capacitor." "Capacitor" leans toward "1.21 gigawatts," "DeLorean," and "Doc Brown," and only a little toward actual electrical engineering, because the internet has written more about the movie than about the physics. That web, and those weights, are the map. That's what the AI is. Not a library. A map of how everything relates to everything else, drawn by counting.

(Engineers reading this are already typing. Yes. It's a simplification. It's the right one.)

Now look at what you usually hand it.

A document is a story. Top to bottom, about one thing. Your employee handbook is a document. So is the email you forward to every new hire with the subject line "READ THIS." When you paste a document into the AI, you're handing it a story and asking it to dig the relationships out. It can. It's just spending its effort on the digging instead of the following, and it's not great at digging.

A knowledge graph skips the story. It's a list of facts and, next to each fact, what it connects to. This rule relates to that one. This kind of instruction requires that kind of follow-up. This thing is never allowed near that thing. Places, and the roads between them. Pure relations, no filler. Which is to say: the same shape as the thing in the AI's head. You're not giving it something to read. You're speaking its language.

Here's what a few entries in mine actually look like. This is copied straight out of the file:

"Active voice rule": "Use active voice to make clear who's performing the action"
"Sequential steps formatting": "Format sequential steps as numbered lists"
"Condition placement": "Place conditions before instructions, not after"
"actionable error message": "After explaining the cause of the problem, explain how to fix the problem."
"heuristic: bath-test": "If you can imagine reading something in the bath, probably, it's explanation."

A few hundred of those. One per line. No chapters. And yes, one of them is called the bath test.

So here's what feeding the AI a graph actually does. It puts your thumb on the scale.

The internet gave "error message" a weak connection to "here's how to fix it," because most of the internet's error messages don't bother. My map made that connection heavy. Not by arguing. Not by explaining. By stating it, in the same format the AI's own map is in, and leaving it in front of the AI while it writes. Every relation in the graph is a little extra weight on one side of a scale the internet already loaded. A few hundred of them, and the scale tips.

Three things make this work better than a handbook, and then we'll get back to the time machine.

Relations are the native format. In the map, "error message" is welded to "how to fix it." There is no road from one that doesn't lead to the other. Unless you're going back to 2015, then where you're going you don't need graph prompts.

So when the AI writes the part about things going wrong, the fix comes along, because in the world the map describes, an error without a fix isn't a thing that exists. A handbook can say the same thing on page 40. But page 40 is a long way from page 3, and the AI, like your employees, is mostly on page 3.

It's flat. One relation per line, every line at the same level. The AI reads a flat list the way you read a menu: all of it, at roughly the same weight. It reads a nested handbook the way you read a nested handbook, which is to say it gets to Chapter 4, Section 2, Subsection (b) and has forgotten what Chapter 4 was about. So have you. So has everyone. Flat means every thumb lands on the scale.

You build it once. The map doesn't know anything about flux capacitors. It doesn't know about time machines. It knows what good instructions look like. I could hand it a coffee maker tomorrow and get a coffee maker manual. In a real system you don't even paste the whole thing in; the agent walks the map, sees how the roads connect and why, and adjusts its mental pattern of what's true based on the topography of information. Its the graph's "shape" rather than content that helps the agent most. For an afternoon experiment, paste the whole map. It works.

That's it. A knowledge graph is a map of what relates to what. So is the AI. Hand it one and it can't help but follow it. The name is doing a lot of work to make a simple idea sound like it needs a consultant.

What came back

The first draft, no map, was a sixteen-section engineering specification, and I want to be fair to it: it was not bad. It was organized. It had a document ID. It had a revision number, which for a document about a fictional car part is a level of ambition I respect.

It also reminded me the flux capacitor was fictional seventeen times. I counted. I'm not proud of it. It opened with this:

Important: The operating principles and specifications in this document are fictional. The Flux Capacitor does not exist as a real time-travel technology.

And then kept going. "Fictional temporal field." "Fictional interlocks." "Routine fictional maintenance." At one point the parts list included a "fictional flux-field emitter," as opposed, presumably, to the real ones. It read like a manual written by someone who was worried I'd sue.

It drew me four diagrams out of keyboard characters. This is one of them, in full:

0 mph
  |
  V
40 mph
  |
  V
70 mph
  |
  V
87 mph
  |
  V
88 mph
  |
  +----> TEMPORAL TRANSITION

That's not a diagram. That's counting.

The second draft, with the map, was a different kind of document. Not a spec. A manual. The kind you'd hand to the person who actually has to drive the thing. The fictional disclaimer was one line, at the top, in a note, and then it mostly got on with its life.

But I promised you proof, not adjectives. So here are three receipts. For each one: the rule from the map, what the AI wrote without it, and what it wrote with it.

Receipt one: who's doing what

The rule:

"Sequential steps formatting": "Format sequential steps as numbered lists"
"imperative-start": "Start all items in a numbered list with an imperative verb (a command such as open or start)."

Without the map:

Phase 1 — Destination Programming The operator enters the desired destination date and time into the time circuits.

Phase 2 — Capacitor Energization Energy is routed into the Flux Capacitor.

Phase 3 — Vehicle Acceleration The DeLorean accelerates toward the required transition velocity.

With the map:

To initiate temporal displacement:Start the DeLorean and enable the time circuits.Enter the required destination date and time.Verify that the destination display contains the correct temporal coordinate.Ensure that the high-energy power source is available.Accelerate the vehicle toward 88 mph.Maintain a stable trajectory during acceleration.Allow the temporal displacement controller to trigger when the vehicle reaches 88 mph.

"Energy is routed." Routed by whom? The first draft doesn't say. It's a passive-voice time machine. The second draft gives you seven verbs and a job.

Receipt two: what to do when it breaks

The rule:

"actionable error message": "After explaining the cause of the problem, explain how to fix the problem."

Without the map:

FC-01 — Insufficient Power Symptom: Flux Capacitor fails to reach full illumination. Cause: Available power below the fictional 1.21 GW requirement. Result: No temporal displacement.

FC-05 — Power-System Failure During Acceleration Symptom: Flux illumination suddenly collapses. Cause: Loss of primary energy source. Result: Temporal displacement aborted.

With the map:

"Result: Temporal displacement aborted." Okay. And then what? The first draft told me what would happen and left. The second draft added a column. That column is the entire difference between a document and a manual.

(Also, that last row is the plot of the first movie, and the manual's advice is "obtain a compatible high-energy source," which, to be fair, is what Doc did.)

Receipt three: the physics gets its own room

The rule:

"key principles of a how-to guide": "maintain focus on the goal — no explaining or reference material, link instead"
"heuristic: bath-test": "If you can imagine reading something in the bath, probably, it's explanation."

Without the map, the theory showed up wherever the document happened to be standing. In the middle of the parts list:

The three-channel geometry provides symmetrical temporal-field distribution and acts as the focal point for fictional spacetime distortion.

Then again as an equation in section twelve. Then again in a closing note about real-world physics.

With the map, it got one section, number fourteen of fifteen, with a sign on the door that said "Theory of Operation":

In engineering terms, the flux capacitor acts less like an engine and more like a temporal field generator. The time circuits specify when the vehicle should arrive, the power system supplies the required energy, and the vehicle supplies the required velocity.

You can read that in the bath. Nobody trying to hit 88 has to.

And two things nobody asked for

The second draft added a safety section on its own. It includes this:

Minimize unnecessary contact with people, objects, or events in a destination period.

Which is the entire plot of the movie, reduced to a bullet point.

And it ended like this:

Without the flux capacitor, the DeLorean remains an automobile. With it, where it's going, it doesn't need roads.

The first draft ended with "END OF DOCUMENT." The one with the map was also the one that had time for a joke. I don't fully know what to do with that.

Why it worked

You already know. But here it is in one breath, because you'll have to explain it to someone.

The AI writes by predicting the most likely next word, and "most likely" is decided by the weights in its model. Think of the weights as roads on a map: every two ideas the internet put together got a road, and the more often, the wider. With only a one-sentence request in front of it, the AI takes the widest roads out of town, and from "flux capacitor" every wide road goes to movie trivia. So you got a fan wiki in a lab coat.

Put a graph in front of it and the roads don't move. The route does. The graph sits in front of the AI while it writes, and it checks it before every word, so the roads that go where the graph points become the ones it takes. Nothing is enforced. No switch gets flipped. When the graph welds "error" to "fix," a fix becomes the most likely thing to write after an error. When the graph says theory lives in its own section, "and now, some physics" becomes a less likely thing to write in the middle of step four. The AI didn't get smarter. You put a thumb on the scale.

That's the whole mechanism, and it's why your overdue-invoice email came back generic. The internet's map has a twelve-lane highway to "polite overdue-invoice email" and a gravel path to yours. Nobody drew the route.

The part where I'm honest

This was one experiment. One request, sent twice. I did not run it fifty times and average the results. I did not have a panel of judges. I did ask another AI to pick the better draft, and it picked the one with the graph, but it also knew which one was which, so that's a little like asking your dog which of two people it prefers and being told it's the one holding the ham. If you put this in a slide deck as "a study," I will find out.

But the difference wasn't subtle. You just read it. The mechanism isn't controversial. And the experiment costs you nothing to run on your own business, this week.

Your afternoon

Write down the rules. Not the mission statement. The rules. The stuff you'd say to a new hire in their first week and then repeat, with slightly less patience, in their second. We never say "reach out." Every email ends with the next step. Never quote a price without the setup fee. We call it a "session," not an "appointment." You have fifty of these. You've just never seen them in one place.

I've found the easiest way to do this is to utilize your commute home trauma dumping to Chat-GPT's voice assistant about your day. After a week it'll have enough context to draw up a knowledge graph about what's important about your job (it knows what graphs are and how to draw connections). Just start by telling it what you're doing and to not interrupt you and you'll be off to the races.

Pro Tip: Keep it flat. When you have the AI write your graph you have to keep it on a short leash. Left to its own devices, the AI will try to write a corporate handbook. It loves "Chapter 4, Section 2, Subsection (b)(iii)." Don't let it. Tell the AI: "Turn my notes into JSON. Keep it flat—two levels deep at an absolute maximum." (trust me ... even when working to refine this article I had to pull my editor back).

In a graph, "flat" just means no folders inside of folders inside of folders. You get a category, the rule inside it, and that's it. Any deeper than two levels, and the AI loses track of the rules just like your employees do. And don't let "JSON" intimidate you. It sounds like something you need an IT guy for, but it’s literally just a text file where everything has quotation marks and colons around it (you saw what it looks like earlier in this post). A flat JSON list skips the filler and feeds the AI exactly what it wants to eat. You just built a knowledge graph. Nobody needs to know it was that simple.

Paste it in before your request. Every time. That's the map. That's the trick. There's an entire industry of people who'd like you to believe it's more complicated than that, and I am, technically, one of them.

One last thing, because it's the actual finding.

If you sit down to write the list and realize you can't, that the rules live in your head and one senior person's head and nowhere else, the AI didn't fail you. It just noticed you don't have a map. It's been noticing for a while. It was too polite to say.

Where you're going

Here's what I keep coming back to.

The AI already has a map of the whole world, and it's a good one. It's why the first draft had a document ID, a revision number, and a sincere diagram of the number 88. Nobody taught it those. It picked them up from the highways.

What it doesn't have, and will never guess, is the dozen roads that make your organization yours. That a complaint gets one apology and then a fix. That you say "session," not "appointment." That the invoice email sounds like you and not like a bank. Those roads exist in exactly one place, and there's no bigger model coming that will find them, because they aren't on the internet. They're in your head. The AI can only take the highway. You have to draw the exit.

So that's the whole post. A one-line request, sent twice. The difference was a map. And the map is the one thing in this story you can't buy, download, or wait for. It's the one thing only you have.

Draw it.

Where you're going, you're going to need roads.