Chat GPT Prompts You Can Reuse, With the Slots Shown

The top result for chat gpt prompts, at least the day I looked, is a Reddit thread by somebody who saves every good prompt they find and posted fifteen of them. That's the real search intent right there. People aren't looking for a lecture on prompt engineering. They want something they can keep.

Fine. So here's the honest version of that. A saved prompt only survives if it's a template with the variable parts marked, because the part that made it work for the person who wrote it is the part that's specific to them, and that's exactly the part that won't apply to you. Mark those bits as slots you fill each time. Keep the rest fixed. A template with four labelled slots will outlive a hundred finished prompts, and it takes about ninety seconds to convert one into the other.

Anthropic's own prompt library does this, incidentally, which is how I know it isn't just my hobbyhorse.

Prompt Versus Template, And Why It Matters

A prompt is a thing you sent once. A template is a thing you can send again.

That sounds like semantics until you look at what happens to a copied prompt in practice. Say the prompt is "act as a senior marketing strategist and write a 30 day content plan for a SaaS company targeting HR managers". Great prompt, for somebody selling to HR managers. You aren't. So you edit it, and you edit the parts you notice, which are the obvious nouns, and you leave the parts you don't notice, which are the assumptions baked into "30 day" and "content plan" and "SaaS". Now you've got a prompt shaped like somebody else's business.

The template version marks what's variable up front.

Act as a {role}. Write a {timeframe} {deliverable} for {business}, targeting {audience}. Constraints: {constraints}.

Uglier. Also honest about where the thinking has to happen. When I look at the official Claude Code prompt library on Anthropic's docs site, that's exactly the shape they publish in, prompts written with braces around the variable pieces and an example fill sitting next to each one, things like "explain what {path} does and how data flows through it, write it up as {format}". If the people who build the model publish templates rather than finished prompts, that's a reasonable signal.

The Five I've Actually Kept

I want to be upfront that I don't have forty of these. I have five that survived a year, and a graveyard of clever ones I used twice.

Two of those have a "don't do the thing yet" instruction in them, and that's not accidental. The model's default is to be helpful immediately, so if you want diagnosis rather than a solution you have to say so explicitly, or you'll get a fixed version of your draft instead of a list of what's broken in it.

The one about asking me three questions is the one I'd hand somebody first. It works because the failure in most writing tasks isn't the writing, it's that a decision hasn't been made, and being asked directly surfaces that faster than any amount of drafting.

One Template, Three Fills

Take the review template from the last row and watch what the role slot does. Same artifact, same purpose, three different fills.

Fill one, role is "a copy editor". What comes back is line-level. Repetition, tense drift, a paragraph that's doing two jobs. Useful, narrow, and it never questions whether the piece should exist.

Fill two, role is "a sceptical reader who has read forty articles on this topic". Now the objections are about novelty. What's here that isn't everywhere else, why should anyone finish this, which claim is the one they'd doubt. Much more uncomfortable and much more useful.

Fill three, role is "the person this piece is trying to sell to". Different again. Objections become practical, along the lines of how much this costs me, what happens if it doesn't work, why should I trust the person writing this.

Same template. Three genuinely different reviews, and none of them would have come from asking "give me feedback on this". That's what a slot buys you. It's not that the prompt got smarter, it's that you got a dial.

Worth noting what didn't change much. Swapping the purpose slot between "for publication" and "for accuracy" moved things less than I expected, and I'd guess that's because purpose is doing work the role slot is already doing. Slots that overlap are wasted slots, and I've since cut two from templates I was carrying around.

The Ones That Died

A mega-prompt that assigned four experts to argue with each other. A "think step by step, then critique your own answer, then revise" chain. Something with a scoring rubric in it where the model rated its own output out of ten before rewriting.

All of those produced impressive-looking sessions and none of them survived contact with actually needing something done. The panel of experts mostly agreed with itself. The self-critique step produced critique that was correct and generic, and the revision was worse about half the time. The scoring rubric gave everything a 7 and then an 8.

I think the pattern is that elaborate prompts are fun to build and they optimise for the wrong thing, which is the feeling of having a sophisticated setup. The five that stuck are all short, all single-purpose, and all boring. That ratio, five kept out of maybe twenty-five tried, is probably the realistic hit rate and it's worth knowing before you start, so you don't conclude you're bad at this when a clever one flops.

There's also a category that died from model drift rather than being bad. Prompts that carried workarounds for things the model used to do wrong. They still work, technically, they're just carrying instructions for a problem that no longer exists, and every one of those lines is spending your attention for nothing. Worth rereading your saved ones once in a while with that specifically in mind.

What Should Never Live In A Saved Prompt

The main way saved prompts go bad is that people save too much of them.

The right column is context, and context is the slot that makes prompts work. It's also the slot that can't be saved, by definition, which is the thing the hundred-prompt posts can never fix no matter how good their prompts are. I went into that at more length in the chatgpt prompts guide.

There's a middle category too. Constraints that you'll always want but only discovered by being annoyed. My list has "no rhetorical questions", "no three-item lists", "don't restate the question before answering". Each of those went in after the fourth or fifth time I deleted the same thing by hand. Building that list is slow and it's the highest-value thing in my whole setup, because it removes work permanently rather than per-task.

Where To Actually Keep Them

Not in chat history. I lost a good one that way, scrolled past it, couldn't find it again, rebuilt it worse.

A plain text file works. So does a note app. The custom instructions field, if the tool you use has one, is the right home for the standing constraints specifically, since those apply to everything and you'd otherwise be pasting them constantly. Project or workspace features, where a set of instructions attaches to a group of chats, are better again for anything role-specific.

What matters more than the storage is keeping a note next to each template about what you changed to make it work. Six months on, the template might need updating for whatever the model does now, but "make it list problems before fixing them" is a durable lesson and the exact wording isn't. There's a whole genre of tooling around this now, some of it worthwhile, covered in prompt engineering tools. For most people a text file is genuinely enough.

If you'd rather browse than build, the ai prompt library piece looks at what the big collections are good and bad at.

Image Templates Work The Same Way, Only More So

The template idea pays off harder on pictures than on text, and almost nobody applies it there.

Because a photo prompt has a style half and a subject half, and the style half genuinely never changes across a set. Lighting, lens, film stock, colour treatment, aspect ratio, exclusions, all of that is a fixed block you paste identically. The subject and setting are your slots. Doing it that way is the difference between six images that look like a set and six images that look like six people made them.

I found this out the annoying way, generating covers for a series of books and getting a different visual world each time because I kept retyping the style description slightly differently. "Warm side light" one day, "warm light from the side" the next. That's enough drift to matter. Now the style block gets pasted byte for byte and only the subject line moves.

Since the marginal cost of a generation on my own machine is basically electricity, I can afford to test whether a fixed block is holding, which is mostly a matter of rendering the same block with three different subjects and seeing whether they'd sit next to each other on a shelf.

Questions I Get

How many prompts should I save? Fewer than you think. I've got five, and I'd say anyone with forty saved prompts has a collection rather than a system, since you can't remember forty and you'll default to typing fresh anyway.

Are paid prompt packs worth it? For text, I wouldn't. You're buying finished prompts, which is the problem this whole piece is about, and the free collections are enormous. For images there's more of a case, not because the prompts are better but because the visual vocabulary is genuinely hard to acquire otherwise.

Does saying "act as an expert" help? A bit, and less than people think. What it mainly does is set vocabulary. If you want better reasoning you get more from asking for the objections, or the steps, or the things that could go wrong, than from any amount of describing how senior the imaginary expert is.

Should I write prompts in Markdown or XML? Depends where you're sending them. Anthropic's docs recommend XML tags for structuring Claude prompts, and OpenAI's developer documentation mentions both Markdown and XML for marking logical boundaries in a prompt. So the underlying advice is the same, which is to make the sections visibly separate, and the syntax is worth adjusting per model.

What about really long prompts with lots of pasted material? Put the pasted material at the top and your question at the bottom. Anthropic documents this specifically for long context work and claims queries at the end can improve response quality by up to 30 percent in their testing. I can't verify their figure, but the ordering is free to adopt and it's the opposite of what most people do instinctively.

What I'd Do This Week

Take the three prompts you've retyped most often. Find the parts that changed between uses and put braces around them. That's the whole conversion.

Then start the annoyance list. Every time you delete the same thing out of an output by hand, write it down as a constraint. Mine took months to build and it's the piece of this I'd protect first.

Deeper technique, if you want it, is in chatgpt prompt engineering, and worked examples with the reasoning attached are in ai prompt examples.