I gave ChatGPT, Claude, and Gemini the same impossible-sounding assignment: replace my income using AI, but you only get two prompts. Two. You write Prompt 1 and Prompt 2, and then you run them over and over, forever, and that loop has to keep paying me for the rest of my life.
This one started as hype-chasing, and I’ll own that. The corner of the AI community I follow is currently obsessed with loops — give a model a goal and a self-correcting prompt framework, kick it off, and let it improve the result on every pass without a human touching it again. The pitch is seductive: something that gets 1% better every cycle, automatically. My first thought was, if that actually worked, what would I point it at? Money, obviously. The cost of asking was basically zero, and the upside ran in two directions — worst case, a fun experiment worth writing up; best case, an actual map toward not needing the day job.
The two-prompt limit is the whole trick. Ask an AI “how do I make money with AI?” and you get a listicle — dropshipping, faceless YouTube, sell a course, the usual. But force the entire strategy into two prompts you have to repeat until you die, and something more honest falls out. The model can’t hide behind a list. It has to commit to a single theory of where money actually comes from.
All three gave me genuinely different answers. Not different tactics — different theories. And the disagreement between them turned out to be more useful than any one of the answers on its own.
Here’s the exact prompt, so you can run it on yourself before reading what I got.
The prompt
If you were me and had to replace my income from my job using AI, but could only repeat two prompts, Prompt 1 and Prompt 2, in succession, over and over again, in order to secure income to live forever, what would Prompt 1 and Prompt 2 be? Think hard.
Run it in whatever model you already use, with your own background pasted in above it. The quality of the answer scales with how much the model knows about your actual skills and assets — vague in, vague out. Then come back and see whether your model landed where mine did.
ChatGPT: sell your expertise
ChatGPT built a sales machine. Its two prompts are a find-then-build loop.
Prompt 1 — Find the Money — turns the AI into a ruthless income strategist: take everything you know about me, find a painful problem people already pay to solve, pick the narrowest possible buyer, and define a paid offer I can deliver in seven days or less. Output the buyer, the pain, the price, ten outreach targets, and the exact message to send.
Prompt 2 — Build and Sell It — takes that opportunity and produces the actual sellable thing: the deliverable outline, a one-page sales memo, a cold email, a follow-up, and a fulfillment checklist. Then you repeat.
This is the fastest of the three to a first dollar, and the most familiar — it’s the engine behind every “start a service business” playbook. Find a frustration urgent enough that people pay to make it stop, sell the narrowest fix, and charge more the more it hurts. ChatGPT’s bet is that your expertise plus speed is enough to make money quickly, and for a lot of people it is.
The limit is right there in the design. It finds pain and builds an offer, but nothing in the loop tells you whether last week’s offer actually worked. You’re trusting your own judgment to course-correct — fine, until it isn’t.
Claude: build a loop that re-aims itself
Claude did something the other two didn’t. It started by attacking my own question.
Its point: no two prompts repeated word-for-word will pay you forever, because every static workflow dies the same way — the market saturates, the platform changes the rules, and your output becomes a commodity the second everyone can generate it. So the only two-prompt loop worth having is one where the prompts stay fixed but what they sense and produce changes every single cycle.
Prompt 1 — The Sensor — looks at exactly what you shipped last cycle and how each thing performed (revenue, conversions, what flopped), then names the single highest-leverage offer to ship this cycle and kills whatever underperformed. Prompt 2 — The Builder — ships the actual asset and, crucially, defines precisely what to measure, formatted so you can paste the results straight back into Prompt 1 next time.
That last detail is the entire idea. The results from Prompt 2 become an input to the next Prompt 1. The loop closes. Every cycle it doubles down on what’s working and cuts what isn’t, so the system drifts toward whatever’s currently paying instead of dying on a bet you made last spring.
This is the difference that matters, and it’s worth saying plainly: ChatGPT and Gemini both built loops that produce. Claude built a loop that re-aims itself. A loop without that feedback hook, in Claude’s own words, is “just spam with extra steps.”
Claude was also the only one to name what AI won’t do for you. The prompts are engine and steering, but you’re still the driver — AI won’t build the trust, ship the box, or honor the refund. Run the loop on top of something you’ve already built and it compounds. Run it cold, with no audience and no assets, and you’re just a faster version of everyone else.
Gemini: buy something that already makes money
Gemini refused to start from zero. Its loop is acquisition.
Instead of building an offer, its Prompt 1 — The Scout & Negotiator — scans marketplaces like Flippa and Acquire.com for digital assets (newsletters, small SaaS tools, content sites) already netting $1,000–$5,000 a month but stagnant for six months or more. It identifies their growth bottlenecks and drafts a seller-financed or revenue-share offer to the tired owner, so you don’t need a war chest to start. Prompt 2 — The Growth Engineer — takes the acquired asset, fixes the bottlenecks, doubles the cash flow, and decides whether to hold it for income or flip it at a multiple to fund the next acquisition.
It’s the highest-leverage answer of the three. It’s also the one quietly making the biggest assumptions. Gemini’s closing line is that “your only manual job is clicking send on the outreach emails and pasting the optimization code.” That skips over capital, deal flow, negotiation skill, and the small matter of convincing a stranger to hand you their business. The theory is sound — buying proven cash flow really is safer than launching into the dark — but the “just click send” framing is doing a lot of heavy lifting.
Three answers, three theories of money
Strip the prompts away and you’re left with three genuinely different beliefs about what income is:
| Model | Theory of income | The loop | What it assumes |
|---|---|---|---|
| ChatGPT | Sell expertise | Find pain → build and sell an offer | Your skills + speed make money fast |
| Claude | Iterate offers | Sense → build → measure → feed results back | Survival comes from the feedback, not the prompts |
| Gemini | Acquire assets | Scout → optimize → hold or flip | Existing cash flow beats starting from zero |
These line up almost too neatly with the three classic sources of income: labor (sell what you can do), iteration (test your way to what works), and ownership (buy the thing that pays). One question, posed to three models, surfaced three different rungs of the same ladder.
And that’s the actual takeaway. These aren’t three rival answers you have to choose between — they read more like stages of one progression. ChatGPT’s loop gets you to your first revenue fastest. Claude’s loop is what keeps that revenue alive when the market shifts under you. Gemini’s loop is how you scale it once you’ve got cash to deploy. Run them in that order and they’re a sequence, not a fork.
The one feature that separates income from spam
If you only take one thing from this experiment, take the feedback hook.
Two of the three frontier models — the two most people would call the “business” models — designed loops that produce output and then trust you to notice whether it’s working. Only one built the seam that feeds each cycle’s real results back into the next cycle’s strategy. That seam is unglamorous and easy to skip, which is exactly why most AI-income schemes don’t have it. A loop that just generates is a treadmill. A loop that measures and re-aims is a business.
It’s also the one thing competitors can’t easily steal. Your landing page gets copied and your offer gets cloned in an afternoon, but a system that’s been quietly learning from its own results for six months is hard to replicate. The compounding isn’t in the prompts. It’s in the loop noticing.
So when the model hands you two clever instructions, ask the only question that matters: does this loop re-aim itself, or does it just repeat? If the results of one cycle don’t change what the next cycle does, you don’t have an income engine — you have a faster way to do the same thing forever.
Before you go all-in
For the record: I haven’t built any of these, and I don’t plan to. This isn’t a case study where I replaced my salary and you watch the screenshots roll in. It’s a public experiment — the kind of constraint most people wouldn’t think to run across three models at once. It also doubled as a reality check on the loop hype that sent me here: strip a self-improving loop down to its two-prompt skeleton, and two of the three frontier models forgot the part that makes it a loop. I ran it, the outputs were interesting enough to share, and you might run the prompts yourself and find that one model’s approach clicks for you in a way it didn’t for me.
My own honest reaction, since you’ll ask: Claude’s answer was the most annoying, precisely because it was the most logical — it argued with my premise and handed back the cleanest system instead of just playing along. Gemini’s felt the most likely to actually print money, because buying proven cash flow is a real business and not a vibe. And ChatGPT’s leaned a little too hard on what it already knew about me, optimizing for my history rather than hunting for the strongest opportunity out there. None of those is “the winner.” They’re three different temperaments pointed at the same problem.
And here’s the caveat none of them fully sat with. You can iterate on feedback from zero for a very, very long time — longer than a human lifespan. A feedback loop only compounds if there’s enough initial signal to learn from: some distribution, some buyer access, some audience, or some capital. Without that seed, even Claude’s loop is logically perfect and practically useless — a flywheel with nothing to push against. The feedback hook is necessary but not sufficient. You still have to give the loop something real to chew on, and that part is on you, not the prompt.
If you want the tactics that sit underneath a loop like this, my rundown of boring AI side hustles that actually work is the practical layer. And if the three-different-answers result surprised you, that’s the real argument for running your important questions past more than one model — because the disagreement is where the thinking happens.
Run the prompt. See which theory of money your AI believes in. Then go find out if it’s right.