The Half-Life of a Good Prompt

Open the folder where you keep your saved prompts. Scroll past the first dozen. Somewhere around item forty you'll find things that used to be genuinely great — a rewrite instruction that made your emails sound human, a summarizer that nailed the tone of your weekly report, a snippet you copied out of a newsletter in a fit of optimism.

Try one. It probably works okay. Which is the problem. It doesn't fail loudly. It just doesn't do the thing you saved it for anymore, and you don't notice because you never re-read the output closely enough to compare.

That's prompt debt: shortcuts taken in the past that now cost you attention you didn't budget for.

Debt is the right metaphor, and also the wrong one

Technical debt makes sense because it has teeth. Code breaks. The build fails, the tests go red, someone gets paged. A stale dependency announces itself.

Prompts don't get that courtesy. There's no compiler, no linter, no test suite. An outdated prompt returns fluent, confident, plausible output that's subtly misaligned with what you now need — and the only signal is a vague feeling that you're editing more than you used to.

So the debt accumulates interest quietly. You pay it in rewrites.

Three ways a prompt goes stale

The model moved. Models get updated, and updates change defaults: verbosity, formatting instincts, how eagerly they add caveats or bullet points. A prompt that says "be concise" meant one thing eighteen months ago and may mean something noticeably longer now. You didn't change anything. The floor did.

The tool moved. The feature got renamed. The context window grew or shrank. The system prompt underneath your prompt got rewritten by someone you'll never meet. Half the tricks in your library were really tricks about a specific interface, and that interface is a moving target.

The job moved. This is the big one and the least discussed. Most saved prompts are answers to a task that existed on a particular Tuesday. The client with that voice, the report with that structure, the deck that needed that exact framing — those are gone. The prompt survives them like a spare key to an apartment nobody lives in.

There's a fourth, slower rot too: context contamination. Someone pastes in a paragraph of a real project, a name, a link, a house style. Now the prompt is only correct for one situation, and you've forgotten which one.

Why nobody cleans it out

Saving takes one click. Deleting takes a decision. And deleting a prompt feels like throwing away work — even when the work in question was a five-line paragraph you wrote during a meeting in March. Storage costs nothing, so the junk drawer grows, and the good stuff gets harder to find inside it.

The real cost isn't disk space. It's trust. Once you can't scan your library and know which entries are current, you stop using the library and start writing prompts from scratch. The archive becomes decoration.

A quarterly paydown, about 25 minutes

Put it on the calendar once a quarter and do it in one sitting. Four passes, in order:

1. Sort by last used. Anything untouched in ninety days moves to an archive folder. Not deleted — moved. Deleting is a decision; moving is a shrug, and shrugs are what actually get done.
2. Collapse the duplicates. If you have eleven variations of "make this clearer," you have one prompt and ten copies of your own anxiety. Keep the two that genuinely differ, retire the rest.
3. De-personalize what survives. Strip the client name, the pasted paragraph, the link. What's left should be an instruction, not a memory.
4. Date-stamp and re-run. Add a single line at the top: what it was written for, and the month and year you last checked it. Then actually run your keepers on whatever model you're using now and read the output against what you wanted.

That last step is the whole point, and it's the one everyone skips. A prompt library you haven't tested is a library of guesses.

Keep a short shelf, not a warehouse

Here's the opinion part: most people would be better off with six durable instructions than six hundred clever ones.

Durable prompts describe the task, not the model. They specify audience, constraints, output shape, and what should trigger a refusal or a flag — the things that stay true when the underlying system changes. "Give me three options with tradeoffs and don't recommend one unless I ask" survives a model update. "Start your response with 'Certainly!' and then use exactly four bullets" is a bet on a quirk, and quirks are the first thing to go.

Curate a small shelf like that, reread it quarterly, and let everything else be disposable. The one-off tricks were never the asset. Your judgment about what you want was.

Six months from now, the shortcuts will be different anyway. Keep the habit; skip the hoarding.