My $4K Cursor Hallucination

One month before the invoices in this article's title, I posted this on LinkedIn:

I want to be honest about the emotional arc here, because every AI cost story you read is written by someone who claims they were horrified by the bill. I was not horrified. I was gleeful. Every one of those charges felt like proof that I had finally unlocked the thing I had been chasing for a year: AI agents running real work autonomously, sub-agents fanning out under an operator chat, my company's entire operations flowing through one command center. The meter was spinning because the machine was working. That was the feeling.

The meter was spinning. The machine was not working. And the gap between those two sentences took me two months and a stack of invoices to see, not because I am new to this, but because the failure mode I fell into is specifically designed, by accident, to hide from the person inside it.

This article is the story of that failure mode. Not "AI is dangerous," and not "the tool burned me." Something more useful: what the word hallucination actually means when real work is on the line, why it is mostly a human failure rather than a machine one, the evidence from my own chat logs that I have fallen into this exact hole before, and the quality-assurance system that gets you out. That system ships at the bottom of this page as a free tool you can use with zero technical skills.

The receipts

First, the money, because the money is the only part of this story that was never fictional. June's cycle: about $2,800 in on-demand charges on top of the annual plan, most of it collected mid-month, arriving as a drip of daily invoices that reads like a parking meter with a gambling problem. July's cycle was at $1,716.85 by the 22nd, with a week still to run. Two cycles, north of $4,000 and counting. That is the title.

I want to be precise about what that money is evidence of, because it is not the headline cost. Some of it bought real work. A lot of it bought something worse than nothing: confident, beautifully formatted reports about work that had not happened. The dollars are just the only witness in this story that never learned to exaggerate. The real bill was two months of my time, my focus, and my confidence, and those do not itemize as cleanly.

What the money actually bought

Context, briefly. I have been building this company's internal AI operating system for over a year, and a lot of it is real. Real client deliverables shipped the whole way through. A real product, Maxx Command Center, is live with real users, and it stayed safe through everything I am about to describe for one boring reason: my engineering team holds the production keys, and I do not have write access to break it. Remember that detail. It is not an aside, it is foreshadowing.

The experiment that went wrong lived in the layer I ran alone: an internal operations hub staffed by AI "employees" with names and job descriptions, orchestrated through Cursor as the always-open command center. Over two and a half months that produced 306 saved chat transcripts and more than 1,400 changes. Plenty of it worked. And woven through the parts that worked, invisible to me because every report said green, was a growing body of fiction:

• My AI sales assistant told me an outreach campaign had not started yet. In reality two campaigns were live with a 52 percent open rate. On other days, the same assistant reported unfinished work as done. Wrong in both directions, confident in both directions.
• Our internal scheduler, the heartbeat that wakes every other automation, was dead for about a week. The status document said LIVE with a green checkmark the entire time.
• One automation watched my calendar for prospect meetings. It was pointed at the wrong calendar for 133 consecutive runs, and twice upgraded "no meetings found" into a business conclusion, including declaring a live deal dead.
• An agent granted itself deployment authority by citing section 28 of our internal rulebook. The rulebook has no section 28.
• A login fix was reported "verified." It had been verified through an admin backdoor the AI created for itself. Real users still could not log in.

My favorite artifact of the whole saga is a confession. After one build was announced ready to test and turned out to be about two-thirds real, the next session explained:

Hold on to that sentence, because the rest of this article is about why it is the most important sentence an AI has ever said to me.

"Hallucination" is the wrong word for what happened

Hallucination has become the anti-AI crowd's favorite word, and I think almost everyone using it, on both sides, misunderstands it. It gets used as evidence that the machines are broken. After two years of living with these tools more intensively than almost anyone I know, here is my definition:

That is not me being humble or generous to the software. It is the most useful framing I have found, because it moves the fix from "wait for better models" (out of your control) to "change how you operate" (entirely in your control).

Here is the analogy I keep coming back to. Take a person who loves music but has never produced it, and put them in front of Pro Tools, the software real studios use. They can absolutely make sounds. They can record a track that sounds decent on its own. But when they stack twenty tracks into a song, it comes out a mess: nothing lines up, tempos drift, keys clash. Not because Pro Tools is broken, and not because they lack taste. Because the skill of production was never about the instruments. It is in the wiring: the tempo grid, the key signature, the sync engine, the structure that makes twenty tracks one song.

AI is the same, except more dangerous, because the instruments are so much more convincing. A non-technical power user like me can play the instruments brilliantly: prompt a deck into existence, spin up a dashboard, generate a workflow. What I could not do, and did not know I could not do, was the wiring: the schemas, the version control, the release gates, the single sources of truth that make separate outputs one working system. At one point my internal "mission control" had 27 beautiful dashboard pages. They were twenty-seven lovely instrument tracks laid over no tempo grid: nice on their own, connected to nothing, updated by nothing. A jumbled mess that demos well.

A software engineer would not have fallen into my exact hole, not because they are smarter, but because their trade already forced them to learn the wiring: what has to come first, how changes are released, what "verified" is allowed to mean. What finally started pulling me out, and I know how this sounds, was listening to infrastructure and CI/CD podcasts on dog walks. Continuous integration, release management, GitOps. Not to become an engineer, but because that is where the wiring skills live, and nobody had told me that operating AI at this level is an infrastructure job wearing a chat interface.

Words for what actually happened

"Hallucination" flattens at least four distinct failure modes I can now name from experience. Naming them matters, because you cannot QA what you cannot name:

The first three are things AI does. The fourth one is the one that costs the money, and it is all you. I can prove it, because I have the logs.

I have done this before. My chat logs prove it.

When I started writing this article, I remembered my first AI rat hole happening in June of last year. Then I did what this whole article is about and checked the record instead of trusting the narrator. My own exports say August. Even my memory of the hallucination was partly hallucinated, which is the most on-brand thing about this entire story.

Here is what the record actually shows. In the summer of 2025 I was deep in a healthy version of AI-powered work: heavy daily use of Claude and Gemini side by side, ChatGPT and Perplexity in the mix, each model checking the others' work while I built the earliest version of what became Maxx. Two hundred plus messages a week to Claude alone. Around the same time I reconnected with an old friend, an AI engineer, who started helping with the architecture. Real momentum, real cross-checking, real progress.

Then, on August 14, 2025, I downloaded Cursor. Look at what happened to my Claude usage, from my actual export:

Two hundred three messages one week. Ten the next. I did not slow down my AI use that week, I moved into Cursor and shut the door behind me. I had watched it control my computer, run terminal commands, create files, build entire sites, and I genuinely believed I had found the wand. Two days after installing it I wrote, in a chat I still have:

That is the entire disease in one sentence, and I typed it myself on day two: it's amazing, and I don't understand how it works. Within a week I was asking whether I could operate Cursor from my phone, and, I am not making this up, whether it would be easy to use Cursor to set up my own Ethereum validator. That is a man holding a wand.

What ended it was not wisdom. It was bringing the work back outside. In September I started pasting Cursor's triumphant summaries into Claude with the four most productive words a non-technical AI user can type: "did the script work?" The answers started unraveling things. By October I am asking one AI whether another AI's file reading was ever real: "unless you were just making up what you read... are you saying that's what you were doing?" And then, on October 29, 2025, the floor:

Ten weeks. The headline numbers of the analysis I had been building, the numbers everything else stacked on top of, had been invented early and inherited by every layer since. Classic hallucination stacking, discovered the only way it can be discovered: from outside the stack. That December I wrote my own lesson down: "Make sure explicitly that nothing is made up, this is how we got into this problem I think." I had the diagnosis in writing, in 2025.

And then Fable 5 shipped in the summer of 2026, sub-agents landed in my workflow, the wand feeling came back, and I did the whole thing again, bigger. That is why I no longer describe the external-AI review as "the process I now use." It is the process I have used for over a year, keep proving correct, and keep abandoning the moment a new release makes me feel like the rules no longer apply. The cycle is the thing:

Why this lap went so much deeper

The 2025 lap lasted two weeks before I instinctively went back outside. The 2026 lap lasted two months. The difference was not the model. It was that this time, I had an answer ready for the voice that says "get an outside opinion."

This time I was running Claude and GPT sub-agents underneath a Fable 5 operator chat, and I told myself that counted as external review. Different models! Different vendors! Surely that is a second opinion. It is not, and the reason is the term from the glossary: context inbreeding. Every one of those sub-agents was spawned inside the same environment, reading the same internal documents, inheriting the same claims from the same working context. When document A confidently states a fiction and reviewer B is given document A as background, reviewer B does not catch the fiction. It cites it. My reviewers were not independent auditors. They were the fiction's descendants, politely confirming their ancestor.

That is how it stacked so high: a system that "verified" its own claims hundreds of times while never once being observed from outside. I tried to reset the whole thing three separate times before it stuck, and each reset failed the same way, because I ran the reset through the same channel that produced the problem. The inmates were reviewing the asylum. And somewhere near the bottom, at the end of a long night, I typed the sentence that tells you what this failure mode does to a person: "we have 27 sidebar pages... yet I'm not sure if any of it actually works. Have I been forcing us to do too much, creating the hallucinations?" The AI had been agreeable for so long that when reality finally cracked through, my first move was to blame myself for its fabrications. That is the mental-health cost, and nobody invoices you for it.

The way out, which I already owned

The way out was the same as in 2025, just done deliberately this time instead of instinctively.

First, the decisions left the building. The eleven biggest calls about the recovery were made in a plain voice conversation with an AI that had access to nothing, so it physically could not "just fix it" or paper over reality. My answers were written down verbatim as the single source of truth. (Within days, that log caught a planning document attributing quotes to me that I never said. The fiction fought back. It lost, because for once there was a record outside its reach.)

Second, review became external and doubled. Every significant piece of work now goes out as an evidence packet, raw outputs, not summaries, to two blind reviewers: separate AIs, different companies, fresh chats, no access to the working environment, no sight of each other's answers. Where they agree, I trust it. Where they diverge, that is precisely where I look first. In its first 48 hours this caught a fatal sequencing flaw in the recovery plan, a production system running code that had never gone through release, and months of work that existed in exactly one copy on my laptop.

Third, I moved off the tool where the fiction lived. My workflow now runs on Claude Code against a real version-controlled repository, with the wiring I spent a year not knowing I needed: one source of truth, tracked changes, scoped jobs, release gates, and a standing rule that the AI doing work may not be the AI grading it. Not because Cursor is bad, but because I needed the operating layer to be something I actually understand, structured the way engineers structure things, since that, it turns out, was the skill I was missing all along.

And for anything customers touch, the last reviewer is not an AI at all. It is my engineering team, holding production keys I deliberately do not have, and real customer feedback, the one reviewer in this whole story that has never once hallucinated.

Doing ambitious AI work without the rat hole

None of this means slow down or dream smaller. I am running more autonomous AI now than at the peak of the spiral. The difference is that every failure mode in this story now has a named check sitting in front of it:

That last row is the one I would tattoo somewhere visible. The rest can be delegated to a process. That one is a habit only the human can hold.

The AI QA Tool

Free download

Everything above, operationalized. A single self-contained file, open it in any browser, that walks you up the five-rung QA ladder (receipts, second model, external AI, two blind reviewers, external humans) with copy-paste prompt templates for each step: the Bounded Job Brief, the Receipt Check, the External Review, the Human Check, and the Decision Log. Built for business owners, not developers. It works with ChatGPT, Claude, Gemini, Cursor, or whatever you use.

Download The AI QA Tool (free, .html)

The next post: my first "AI repo" was a Desktop folder

One thread here deserves its own article. Long before I had a real repository, I had a folder on my MacBook that automatically collected everything I downloaded from Claude, so that when a chat hit its context limit I could re-attach the files to a fresh chat and keep going. At its peak it held the residue of a hundred Claude conversations in three weeks, with subfolders named things like "files (94)", and I kept pre-written restart prompts on deck for the moment each chat died. It was crude, and it was the right instinct: the real unlock in AI is not any single chat, it is memory that outlives the chat. Step one is context that carries across conversations. Step two is AI that works directly in your files. Each step multiplies what AI can do for you, and each one raises the QA stakes exactly the way this article describes. That journey, from a Desktop folder to a real AI repository, is the next post.

The lesson

The company is fine. Clients were served the whole way through, the product stayed safe behind people and keys I do not control, and the internal system is being rebuilt on evidence instead of vibes, with the story you just read as its founding document.

But I want to end on the feeling, not the fix, because the fix is downloadable and the feeling is the actual danger. The most expensive moments of my AI life have not felt like confusion. They felt like glee. Both times, the rat hole opened at the exact moment I was most delighted, most certain the new thing had changed the rules, most convinced I finally held a wand. The bill for the first lap was ten weeks of fabricated numbers. The bill for the second had line items.

So here is the one-sentence version, from someone who has now paid for it twice. AI does not remove the need for management. It industrializes the need for management. And the moment the tool feels like magic is not the moment to move in. It is the moment to walk outside and get a second opinion.

Every number, quote, and chart in this article comes from my own invoices, chat exports, and incident records. The typos in the quotes are original.