ChatGPT Prompts That Work, and Why Lists Don't

I counted what's on the first page of Google for chatgpt prompts. One post offers a hundred. Another has 33. A third has 25, and the Reddit thread sitting above all of them collects 15 that somebody had been saving. You could copy close to two hundred prompts this afternoon and be no better at this than you were at breakfast.

Here's the short version, for anyone who's skimming and about to leave. A prompt works when it fills six slots, and the slot people skip is almost always context. Role, task, context, format, constraints, examples. Fill those and a mediocre prompt beats a clever one you copied. Skip context and the model writes something generically correct that you then have to rewrite yourself, which is the whole failure mode nobody names.

The second thing nobody names is that ChatGPT makes pictures now, and image prompts follow completely different rules. I went through the first page for this phrase. Of the ones that would open for me, not one touches it.

Where I Got Interested In This

Boring answer. I make a lot of things that need a picture attached.

Book covers for the titles I self-publish, trailer frames, thumbnails, product art for a small pile of toolkits I sell between $29 and $49. Somewhere in the middle of that I stopped treating prompting as a trick and started treating it like a form you fill out, mostly because I got tired of rerolling generations and squinting at what changed. When you run image and video generation locally, which I do on an M4 Pro, the marginal cost of another attempt is basically electricity, so I could afford to be sloppy for a while. That turned out to be the problem. Cheap rerolls let me avoid learning why anything worked.

The slot list below came out of finally sitting down and changing one thing at a time.

The Six Slots Every Working Prompt Has

MIT Sloan's guide, which ranks for the broader ai prompts query, teaches three strategies. Provide context, be specific, build on the conversation. Coursera's page names six components and stops there. Both are fine and both leave out the part I actually needed, which is what specifically goes wrong when a slot is empty. That's the column below.

The one I'd defend hardest is context. Anthropic's own prompting docs make the same point for Claude and give a nice example, where instead of writing "NEVER use ellipses" you write that the response will be read aloud by a text to speech engine, so ellipses won't be pronounced correctly. Same instruction, plus the reason. Their docs say the model generalizes from the explanation, and in my own use that's held up more reliably than any phrasing trick I've tried.

Examples is the slot people skip because it feels like work. It is work. It's also, per Anthropic's documentation, one of the most reliable levers on format and tone, and they recommend three to five of them.

A Rewrite, Start To Finish

Here's a prompt I actually sent, unedited, back when I was writing sales copy for one of the toolkits.

write a product description for my AI video toolkit

What came back was competent and completely useless. Fluent marketing English about unlocking your creative potential, aimed at nobody, roughly 200 words of the kind of thing you scroll past. I didn't blame the model. I'd given it one slot out of six.

The rewrite, with the slots labelled so you can see the seams:

Role. You're a direct response copywriter who writes for solo builders, not agencies. Task. Write a product description for a $49 digital toolkit. Context. The buyer is one person who already makes videos badly and knows it. They've tried three tools, abandoned two. They don't want inspiration, they want the specific pipeline. They found this page from a search, not from a newsletter, so they have no idea who I am. Format. 160 to 200 words. One short opening line, three body paragraphs, no bullet list. Constraints. No exclamation marks, no "unlock", no "empower", no rhetorical questions. Don't claim results I haven't stated. Examples. Match the tone of this paragraph from my existing page. [pasted]

That version needed one round of edits instead of four. The change I'd point at isn't the role line, which is the part everyone puts first and which honestly does the least. It's the sentence about finding the page from a search rather than a newsletter, because that single fact tells the model there's no existing trust to lean on, and the whole opening changes shape as a result.

Slot for slot, that's maybe ninety seconds of typing. I've watched people spend twenty minutes hunting for a better prompt to copy instead.

When It Comes Back Wrong, Change One Thing

Four rounds. That was my old average before I started paying attention, and most of those rounds were me rewriting the whole prompt from scratch because something felt off, which meant I never learned anything from any of them.

The habit that fixed it is embarrassingly simple. Read the bad output, decide which single slot caused it, change only that slot, resend. You get a much better sense of what each field is doing, and you stop carrying superstitions around.

Rough mapping, from my own notes rather than anybody's research. If the output is factually fine but reads like it was written for a stranger, that's context. If it answered a slightly different question than the one you asked, that's the task verb being vague, and swapping "help me with" for "write" or "list" or "critique" usually settles it. If the tone is wrong, that's examples, not role, and I got this backwards for a long time because adding another adjective to the role line feels like the fix. If the length or shape is wrong, that's format, and be numeric about it, because "short" means nothing and "under 90 words" means something. If it keeps doing the specific annoying thing you hate, that's a missing constraint, and you should write the constraint down somewhere permanent because you'll want it again next week.

The one that trips people up is the difference between context and constraints, since both feel like extra information you're bolting on. Context is about the situation the output lands in. Constraints are about the output itself. Telling the model that your reader is a first-time buyer who found you through search is context. Telling it not to use exclamation marks is a constraint. Mixing those two into one wall of text still works, roughly, but you lose the ability to diagnose, and diagnosing is the whole point of changing one thing at a time.

Images work the same way, with one wrinkle. Because generation is stochastic, you have to reroll the same prompt two or three times before you conclude that a word did anything, otherwise you'll credit a change for what was just a different seed. I've fooled myself with this more than once. A prompt tweak looked like a breakthrough, and it was noise.

Why Copy Paste Lists Underperform

My complaint with the hundred-prompt posts isn't that the prompts are bad. Some are good. The complaint is structural.

A prompt in a list has to work for everybody, which means the context slot is empty by definition, because your context isn't the list author's context. So what you're copying is the four easy slots and none of the hard one. Then the output comes back generic and you conclude that AI writing is generic, when what actually happened is you used a form with the important field left blank.

There's a version of the list that does work, and one of the ranking pages half-does it. Anthropic's own prompt library for Claude Code is built out of prompts with variable slots written in, things like "explain what {path} does and how data flows through it, write it up as {format}", with an example fill supplied next to each. That's a template, not a prompt. Templates survive being copied. Finished prompts mostly don't. I go further into building your own in the piece on chat gpt prompts you can reuse, where the whole point is the slots.

Also, and this is smaller, prompt lists rot. Model behaviour shifts under them. A prompt tuned for a model from two years ago carries workarounds for problems that no longer exist, and you inherit the scar tissue without knowing what it was for.

So What About The Pictures

Genuine question I get, usually phrased as "does any of this apply to image prompts". Not really, no. Almost none of it transfers.

A text model is reading your prompt as instructions. An image model is reading it as a description of something that already exists, and it's much closer to a search over a very large space than to a conversation. That difference changes everything about how you write. Instructions like "make it good" or "be creative" do nothing, or worse, they eat tokens that could have carried actual visual information. Roles are useless. "You are a professional photographer" is four words about a person who isn't in the picture.

Google's documentation for Gemini image generation says it plainly enough. Describe the scene rather than listing keywords, and the more specific you are, the more control you have. They ship six named templates, and the photorealistic one names its slots as shot type, subject, setting, lighting, and camera angle or lens. That's a form too. It's just a different form.

What An Image Prompt Needs That A Text Prompt Doesn't

I tested this the slow way, one variable at a time, on the same subject across a few dozen generations. Locally, so the only cost was my afternoon. What I ended up with is less a list of magic words than a map of which words move which part of the frame.

Lighting is the one I'd move to the top if I were rewriting this. Change nothing but the lighting phrase and you get a different photograph, not a differently lit version of the same photograph. Swap "low afternoon side light" for "overcast, flat, diffused" and the composition itself shifts, because the model has learned that those two lighting conditions belong to different kinds of pictures taken by different people for different reasons.

Aspect ratio surprised me the first time. I assumed it cropped. It doesn't. Ask for the same scene at 9:16 and the subject gets recomposed for a vertical frame, with more headroom and a tighter horizontal read. Gemini's docs list ten supported ratios for their current image model, from 21:9 down through 1:1 to 9:16, which tells you they expect people to use it as a composition control rather than an afterthought.

If you want the long version of that table with the photography vocabulary spelled out, it's in how to write an ai photo prompt. The subtraction side, which is its own skill, is in negative prompts.

Video Is Harder And I'm Still Bad At It

Confession first. I've shipped a book trailer, and I still find video prompting the least predictable thing I do.

The extra slot is motion, and motion is where prompts stop behaving like descriptions. You're now specifying two things that fight each other, what's in the frame and what changes over time, and models handle the second one much worse than the first. Camera motion helps. "Slow push in", "handheld, slight drift", "locked off" are all instructions a model can act on, and they're more reliable than describing what the subject does, because subject motion tends to arrive as either nothing or chaos.

The thing I learned the expensive way is to keep clips short and specific rather than long and ambitious. One action per clip. Cut between them later. A four second shot that does one thing correctly is worth more than a ten second shot where the hands go wrong at second six, and hands still go wrong. More on this in ai video prompts.

The Handful I Actually Reuse

Three, if I'm honest, and none of them are clever.

One is a rewrite pass where I paste my own draft and ask for a list of sentences that sound like nobody wrote them, no rewriting, just the list. One is a structure check where I paste an outline and ask what a reader would still be confused about after reading it. The last is the boring one, where I hand over a mess of notes and ask for the three questions that would let it be organised properly.

That's it. The rest I type fresh, because the context changes every time and the context is the part that matters.

Questions I Get About This

Are prompt engineering skills going away as models improve? Partly, and the part that's going is the part that deserved to. Magic phrasing, "take a deep breath", threatening the model, tipping it imaginary money. That stuff was always noise. Specification isn't going anywhere, because no model can guess a fact about your situation that you never told it, and that's what most bad output traces back to.

Should I pay for a prompt library? I wouldn't, not as a beginner. Libraries sell you finished prompts, and finished prompts are the copy paste problem in a nicer wrapper. Free ones are everywhere too. There's a well known Claude prompt collection on GitHub sitting at about 5.4k stars as I write this, and it's the same shape as the paid ones.

Does the same prompt work across ChatGPT, Claude, and Gemini? Roughly, at the level of the six slots. Not at the level of formatting. Anthropic documents XML tags as their structuring recommendation and OpenAI's developer docs mention both Markdown and XML for marking logical boundaries, so the syntax you wrap things in is worth adjusting per model. I wrote up the differences in claude prompts.

How long should a prompt be? Longer than you think for text, shorter than you think for images. Text prompts fail from missing context far more often than from bloat. Image prompts fail from stuffing, where you pile on twelve descriptors and the model averages them into mush. I'd say six to nine strong visual clauses is where my own hit rate peaks, though I'd want to check that properly before I'd swear to it.

What about really long documents? Anthropic's docs have a specific recommendation here that I hadn't seen elsewhere, which is to put the long material at the top and your actual question at the bottom. They say queries at the end can improve response quality by up to 30 percent in their testing, particularly with multiple documents. I can't verify their number, obviously, but the ordering costs nothing to adopt.

What I'd Do With A Blank Chat

Write the context slot first. Before the role, before the task, before anything. Two or three sentences about who this is for and what's already been tried and why you're doing it at all.

Then the task, then the format, then whatever constraints you know you'll otherwise have to fix by hand. Skip examples on the first attempt and add them on the second, once you've seen which way the output drifts.

For pictures, invert it. Subject and lighting first, everything else after, and stop adding words the moment the frame stops improving.

I'd also say this. The reason those hundred-prompt posts rank isn't that they're useful, it's that "give me a list" is what people search when they don't yet know what they're missing. I sat a site at average position 47 in Search Console for months while I was still convinced the problem was elsewhere, so I'm not above being the person who's confidently solving the wrong thing. Prompts were like that for me for about a year. The fix wasn't a better list. It was noticing which field I kept leaving blank.