Your favorite prompt got worse. Here's what usually happened.

There's a particular kind of annoyance that comes from a prompt you trust going soft on you. Same words, same task, and the answer that used to be sharp comes back generic. You didn't change anything. That's the problem. Everything around it changed.

It's worth going through your saved prompts now and then like a pantry. Some are fine. Some have quietly gone stale. These are the things I look for, and what I do about each.

The chat got too long

This one is the easiest to miss, because nothing breaks loudly.

I learned it with images. Say you're making a series of pictures of the same character. The first few look right because you attached reference photos at the start. Twenty messages later the face has drifted into somebody generic, and if you scroll up, the references are still right there in the chat. The model just isn't leaning on them anymore. The fix I settled on is boring: attach the reference again with every single request, and check each result against it before moving on.

Text does the same thing, more quietly. The rules you set in message one ("no jargon," "keep it under 200 words") fade as the conversation fills up with other stuff. When a long chat starts ignoring you, don't argue with it. Open a new one and paste the context back in.

The rules piled up

A prompt that's been "fixed" ten times usually has ten patches on it. Each one made sense the day I added it. Together they fight.

Before (a real-looking pile):
You are an expert copywriter. Write a product description. Be concise.
Be detailed. Use bullet points. Don't use bullet points for the intro.
Make it fun but professional. No exclamation marks! Include keywords.
Keep it under 100 words. Mention all features.

Read that as the AI would. Concise and detailed. Fun and professional. Under 100 words and every feature. It can't win, so it splits the difference and you get mush.

After:
Write a product description for the coffee grinder below.
Why: it's for the product page, and most people skim on their phone, so
lead with the one thing that makes it worth buying. Under 100 words.
Mention only the features a home user would care about.

My rule when I refresh: delete before I add. Most of the time the prompt gets better from losing four lines, not gaining one.

It was written for an older model

A lot of older prompts floating around (public libraries included, ours too) are full of tricks that mattered once. ALL CAPS warnings. "Take a deep breath." Paragraphs telling the model to be careful. The newer models don't need that, and some of it makes them worse, because they take every shouted rule seriously and get rigid.

So when a model updates, re-test the prompts you actually depend on. Not with a made-up test, with two or three real inputs from the last week. If the answers got stiffer, try cutting the old scaffolding and just saying what I want and why.

It doesn't know what you already have

This one cost me a whole build session once. I asked an assistant to make a small tool, it made a nice one, and only afterward did I realize I already had that tool. It existed, it was better, and it was named in half a dozen of my own notes. The assistant didn't know because I never told it to look.

Now any prompt I reuse for building things starts with a line like this:

Before you make anything new, check what already exists in [folder or
list] and tell me if something there already does this.

It feels obvious written down. It wasn't obvious at the time.

Keep a notes file

The single habit that's helped most: when a prompt misses, I write down what went wrong and what I changed. One or two lines. Next time I or the AI picks it up, the correction is already there. Over a few months that file becomes a better prompt than anything I could write from scratch, because every line in it was earned.

If you save your prompts in a NeLocker vault, you can keep the note right next to the prompt it belongs to.

A quick refresh routine

When I pull a prompt out to check it, I go roughly in this order:

• Read it out loud. Anything that contradicts something else goes.
• Is the why in there? If not, add one line saying what it's for.
• Cut anything that was there to work around an old model.
• Run it on two real inputs. Compare with what you'd write yourself.
• Write one line in your notes about what changed.

That's it. Ten minutes, maybe less. If you'd rather start from something fresh, the free prompts are a good base, and how to write a prompt that works the first time covers the what-and-why shape I use for new ones. The longer beginner's guide is there too.

Quick answers

Why does a prompt that used to work stop working? Usually one of three things changed. The model was updated, the chat grew so long that early instructions faded, or the prompt collected extra rules over time that now pull against each other.

Should I start a new chat when the AI starts ignoring instructions? Often, yes. In a very long chat the early instructions and reference material lose weight. Start fresh and paste the key context and any reference files again.

How do I keep my prompts up to date? Keep them in one place, write down what you changed and why each time one misses, and re-test your important prompts after a model update with two or three real inputs.