Prompts Like Bread Go Stale.

You have been told the trick is finding better prompts. Buy the list. Paste the prompt. Get the result.

Maybe you bought one. Maybe you bought three.

That advice is wrong twice over.

A prompt is not good or bad on its own. It is good on one machine. And prompts go stale, like bread, because the machine gets swapped out under the same name while you sleep.

God of Prompt sells prompt bundles for a living. We opened its pricing page on Sept 21 expecting marketing.

Found a confession.

Works with ChatGPT, Claude, Gemini, Midjourney and more

Same page, further down:

when a model changes, the affected prompts and guides are reworked and land in your account

Both cannot be true. If one prompt really worked on four machines, a model change could not break it.

Sold as universal. Insured against being universal. Nobody blinking.

The research nobody quotes

UC Santa Cruz and Accenture gave it a name. Model Drifting, meaning a prompt tuned for one machine underperforming on a different one. They took a prompt that scored near perfect where it was built. Changed not one word. Ran it somewhere else.

It fell off a cliff.

Microsoft and MIT came at it sideways and changed nothing but the formatting. Plain text. Markdown. YAML. JSON. Same question every time. Quality moved enormously on wording that never changed, and the wrapper that won on one model lost on the next family over.

Both papers cover older models. We are not pretending a 2024 study measured the machine on your desk. It measured the principle.

Then the one that should stop you cold. About 1,900 people were moved from an older image model to a newer one. Everybody improved. Half the improvement came from the model. The other half came from the people, quietly rewriting their prompts to suit it.

Read that twice if you have ever paid to upgrade.

Half of what you thought you bought, you typed yourself.

Prompts go off

The papers measure carrying a prompt between machines. Nobody has studied what happens when the machine changes underneath you and keeps its name. So this half comes from people who wrote it down this month, right after their tools updated.

Fable 5.1 is the first Anthropic model that just flat out ignored basic instructions (Sept 6)

Opus 5 keeps doing extra stuff that was never asked for (Sept 4)

Nothing on their invoice changed. Nothing in their prompt changed. Everything else did.

A prompt has a best-before date and nobody prints it on the packet.

Same prompt, two machines

That is the lab. Here is the parking lot.

We searched every first-hand public account we hold for the words "same prompt." Most of the hits are somebody running one prompt against two different machines.

Every last one came back different. Not one said it came out the same.

One person's assistant went after his own computer. He fed the identical prompt to ChatGPT, which told him to tick a checkbox. That was the actual problem (Aug 31).

Who is doing the counting

Search this yourself and you will get numbers. Big ones. Check where they come from.

The biggest traces to a blog run by a company selling AI products, and the page would not open for us. The second comes from a company selling prompt services. That is not research, it is a brochure with a decimal point.

The prompt industry is not hiding the answer. It is just never the one holding the stopwatch.

What it is costing you

Three ways, and the last one is the expensive one.

• You blame yourself. The list promised it works, so the only variable left is you. It is not you. It is a prompt shaped for a machine you are not using
• You pay for the retries. Every reshoot burns plan allowance. People are hitting weekly caps in a day
• You cancel the wrong tool. The reads are full of people deciding a tool got dumber, when what happened is they handed it instructions written for a different machine. Then they pay to move

Directions to the same address read differently depending on whether the driver knows the town. And streets get renamed.

Nobody credits a list

We went looking for the genre's own words. Prompt library. Prompt pack. Prompt template. Best prompts for.

Out of every first-hand read we hold, that language shows up almost nowhere. One hit is a guy who built a prompt library for accents. One is a product blurb with the word "infinite" in it.

The word prompt is everywhere. Just never the way the lists assume.

i spent way too many nights trying to prompt-engineer away weird formatting quirks before realizing it's just faster to fix them myself (Aug 13)

I kept tweaking my prompts over and over until I finally got something that wasn't completely off the wall. Could've probably just looked up some recipes online in a fraction of the time (May 2025)

An evening lost arguing with a machine about kale smoothies. Cookbook was right there.

The people who are good at this already aim. Some use one machine to write for another.

I regularly ask it to refine prompts for a local LLM to follow (Aug 15)

And the one that points at the fix. Somebody stopped writing longer prompts and started handing over their own paperwork.

Instead of having to write huge prompts to tell it all the detail it needs to know to solve a problem, I now just download our documentation and drop multiple PDF files into the Gemini chat (April 2025)

You never had a prompt problem. You had a nobody-told-it-anything problem, pointed at the wrong machine.

We built this, and it is free

Fair warning, this part is about us. Skip it. The steps below work either way.

A separate version of every prompt for Claude, ChatGPT, Gemini and Copilot. Not one prompt with four logos under it. Every live version re-checked this month. The ones that stopped working got retired and replaced, each with a successor behind it.

That last part is the only reason we can say prompts go off with a straight face. We watched it happen to ours.

No bundle, no lifetime deal, nothing waiting at the end of it.

Now hold us to our own test. Which machine was it written for, and when was it last looked at? If those answers ever go missing from our page, we have earned exactly what we just gave that roundup.

The prompt page is here.

Do this Monday

Free, in order.

• Name your machine and date it. Tool, model version, today's date, at the top of the prompt. That date is the best-before
• Write the job down once. One page. What you sell. Who buys it. What you charge and what you will not do for money. How you talk to a customer who is annoyed. One sitting, good for a year
• Paste that page first, then ask. Every time. It does the job a prompt list pretends to do, except it is about you
• Run your best prompt on a second machine. Most have a free tier. Whatever you believe about universal prompts will not survive the comparison
• Re-test after any version change. Ten minutes. Cheaper than a week of arguing with a machine

What to look out for

• "Works with ChatGPT, Claude, Gemini and more" is the warning label. Printed as a feature. It means nobody aimed it
• The lifetime update promise is a confession. If prompts need reworking when a model changes, they were never universal and they do not keep
• A new version under the same name is a new machine. Your invoice will not mention it. Your results will
• A percentage with no paper behind it. Ask who ran it, on which models, and whether you can open the method. Missing any of the three, it is marketing
• The restamp. GoDaddy's library was written May 2024, restamped September 2025, and ends at GoDaddy's own website builder. Sage's was April 2024, restamped July 2026, with Sage Copilot in the callouts. Look for a published date next to the modified date
• The roundup written by a contestant. God of Prompt publishes "Best Prompt Libraries for Small Business Owners in 2026." God of Prompt sells prompt libraries. God of Prompt ranks first (opened Sept 21)
• A price in an article is not a price

• The tracker you cannot verify. Peec and Profound each sit on a single first-hand read in our pool. Not a verdict, and we will not pretend otherwise. A practitioner on Aug 3: "the numbers swing a lot between runs and personalization messes with it. treat them as trend, not truth" (Reddit)

The spin is the product

Count what it took to check one small corner of one question.

Six. One afternoon.

• A pricing page promising one prompt works on four machines, and promising to repair it when a machine changes
• A roundup ranking its own publisher first, disclosed nowhere
• Prices the store does not charge
• Two libraries written in 2024, wearing 2026 dates, each ending at the publisher's own product
• Every large percentage owned by somebody who sells prompts
• A tool named one thing in our registry and another on the graphics

None of it illegal. Most of it not even unusual. It is what a business looks like when everyone explaining the thing also sells the thing.

Which is the whole reason unbought.ai exists. Not to tell you whether AI is good or bad, because nobody can answer that for your business but you. Separate what is known from what is being sold.

So we open the page instead of quoting the summary. We print the date. And when a claim does not survive, it goes up anyway, in the column marked no.

Four claims came in. Two survived. We printed all four.

Where that leaves you

Prompting is a real skill and it is yours. It moves with you between machines and between versions, and nobody can invoice you for it.

What does not travel is a stranger's finished answer, written for a machine they never named, sold as working everywhere and forever.

Prompts are not magic and they are not a con. They are a tool with an address and a shelf life. That is the whole difference between AI fact and AI fiction, and it is the one line nobody selling you a bundle will ever draw for you.

If somebody you know just bought a prompt pack, send them this before they open it. Ask which model it was written for, and when.