Hacks: The 12-Word Prompt That Fixes Most Error Messages Forever

A product manager I know sent me a screenshot last month. It was the error state in his company's billing dashboard. The exact text, every character: "Sorry, an unexpected error occurred. Please try again later."

Under it, he wrote, "This is on 340,000 screens per month. I've been asking for a rewrite since 2023."

Three hundred and forty thousand screens a month. One sentence. Eight words. Zero information. A small daily donation to user rage, collected at scale.

Here is the thing that nobody wants to admit: most error messages in most products were written by the engineer who wrote the function, in the three seconds between writing the catch block and getting back to the actual problem. They are not bad because the engineer is bad. They are bad because "write a good error message" is not the job the engineer is doing, and the content designer who does have that job never sees the string until it is in production.

You can fix this in an afternoon. You do not need a design sprint. You need a prompt.

The prompt

Rewrite this error message. Tell the user what happened, why, and what to try next, in plain language, in that order.

That's it. Twelve words (or thereabouts — I count twelve, but punctuation pedants will argue). Paste it at the top of any AI chat, paste the error text underneath, and ship the output. That is the hack.

I know. This is the part where you think, "there must be more to it." There is more to it, but only a little. The twelve words do the heavy lifting, and the rest is just getting out of the way.

Why it works

The prompt does three things, and each one undoes a specific thing engineers do wrong when they write errors under time pressure.

"What happened" forces the first sentence to describe the actual event from the user's point of view. Not "error code 503." Not "failed request." Not "invalid state." "We couldn't save your changes." "Your card was declined by the bank." "This file is bigger than we can handle right now." An engineer, left to their own devices, describes what the code did wrong. The user does not care what the code did. The user cares what happened to their thing.

"Why" forces the second sentence to explain the cause, but only the part the user needs to understand to decide what to do next. "Your internet connection dropped while the file was uploading" is useful. "HTTP 504 gateway timeout" is not. The AI is surprisingly good at this distinction if you ask for "plain language," because it has read ten thousand bad error messages and knows, at some level, which ones are which.

"What to try next" is the part that makes users love you, because it is the part most products skip entirely. "Please try again later" is the universal cop-out, and it is in the screenshot above because it is the thing an engineer writes when they do not know what the user should try. The prompt replaces it with a real suggestion: "Check your connection and try again," or "If this keeps happening, contact support at the email below," or "Your card issuer may have blocked the charge — call the number on the back of your card."

Notice what the prompt does not ask for. It does not ask for apology ("we're sorry"), because "sorry" is the single most overused and least valuable word in product copy. It does not ask for empathy ("we know this is frustrating"), because the user is a grown-up and does not need you to narrate their feelings back at them. It does not ask for brand voice, because brand voice in an error message is like a waiter doing stand-up while your food is on fire.

The only thing the user wants from an error message is information and a next step. The prompt asks for exactly that, in exactly that order, and stops.

A real before-and-after

Let's run the product manager's error through the prompt.

Before: Sorry, an unexpected error occurred. Please try again later.

After (from the prompt, with no additional context): We couldn't load your billing details. This usually means our billing provider is temporarily unavailable. Wait a minute and refresh — if it still doesn't work, email support and we'll look into it.

Two sentences and a clause longer. One hundred percent more useful. Zero apology. It does not use the word "unexpected," which is meaningless (every error is unexpected from the code's point of view; no error is unexpected from the user's, because they're seeing one). It does not use "error occurred," which is the passive-voice fingerprint of a developer writing copy at 2 am. It tells the user what to do and when to stop trying.

This is the kind of rewrite you can generate in under a minute for every error in your product. If your product has fifty errors, that's under an hour. If it has five hundred, that's an afternoon.

The upgrade: context

The twelve-word prompt works on any error in isolation. The upgraded version works better if you also give the AI the context of what the user was doing when the error fired.

Rewrite this error message. Tell the user what happened, why, and what to try next, in plain language, in that order.
User was trying to: [describe the action]
Technical cause: [what actually went wrong]
What we know about the fix: [what the user should try, if anything]

Filling in those three fields takes thirty seconds. The output goes from "good enough" to "actually great" because the AI now knows the user was trying to upload a CSV file, the technical cause was a file size limit, and the fix is to split the file or use the API. The rewrite will say something like: Your file is too big for the web uploader — the limit is 25 MB. Either split it into smaller pieces, or use the API endpoint for larger uploads (link in the docs).

That message is a real message. It will save support hours. It will also make the user feel like a person using software instead of a ticket in a queue.

⚠️prompt-error-message-rewriter is the version of this prompt I keep around for batches. Paste a list of errors, get a list of rewrites. It is designed to be run across a whole product's error catalog in a single sitting. If you've got a CSV of every error string in your system (and your engineers can probably generate one in about ten minutes), you can rewrite them all in an afternoon.

The sanity check

AI can confidently rewrite an error into something that is friendly, clear, and completely wrong about what the error actually means. This is the one place you cannot skip the human review.

The check is quick: read each rewritten message and ask, "If I were a user seeing this, would I take the action it's suggesting? And would that action actually help?" If the answer to both is yes, ship it. If the answer to either is no, you have one of two problems. Either the AI guessed the cause wrong (go back and give it the technical context), or the engineer's original message was concealing a bug that the rewrite exposed. Both are useful findings. The second one happens more often than you'd think.

✏️soul-the-content-design-coach is the soul I'd hand the output to for a second opinion if you have even a hint of doubt. It will catch the tone problems, the "we're sorry"s that snuck back in, and the moments where the AI's rewrite is technically accurate but condescending. ✒️skill-content-design-microcopy-review runs the same pass at a whole-product level if you want a systematic version.

Do this tonight

• Open your product's error log or your frontend codebase. Find the ten most-fired error messages.
• Paste each one, one at a time, into your AI of choice, with the twelve-word prompt on top.
• Spend sixty seconds on each rewrite. Read it, decide if it's accurate, adjust if it's not.
• Ship the fixes.

An hour from now, your most-seen errors will be the best-written strings in your product. Your support team will notice by Friday.

The product manager's dashboard? He ran the prompt on it last month. The new message reads: We couldn't reach our billing provider. This usually clears up in a minute or two — try refreshing. If it's still stuck, email us at [support email] and we'll dig in. Three hundred and forty thousand screens a month, and now every single one of them tells the user something true and useful.

Twelve words. One afternoon. Go.

Part of the a-gnt "Hacks" series.