AI Prompt Examples With the Edits That Fixed Them

Here's what's wrong with nearly every collection of ai prompt examples, including the good ones. They show you the finished prompt. MIT Sloan's guide, which sits at the top for this phrase, names three strategies and gives you a nice contrast between "Tell me about climate change" and a much better version about economic implications in developing countries. The Prompting Guide's examples page runs seven categories, summarisation through reasoning, all solid, all text. One prompt marketplace on the same page advertises over four thousand free ones.

None of them show the edit.

So that's what this is. Each example below is a prompt I sent, what came back, the single thing I changed, and what changed as a result. One lever at a time, because that's the only way you learn which field does what. Three text, three image, one video, plus a ledger at the end and a list of edits that turned out to do nothing.

How To Read These

One change per round. That's the whole method and it's the part people skip, because when output disappoints the instinct is to rewrite everything, and then you've learned nothing except that the second prompt was better than the first.

For images there's an extra step. Generation is random, so run each version at least twice before you decide the edit did anything. I've celebrated a breakthrough that turned out to be a coin flip, and the second render is the only thing standing between you and a belief you'll defend for a year.

Everything below is my own working rather than anyone's study. Sample sizes are small, a handful of runs each, and I'll flag where I'm less sure.

Text Example One, The Decision Behind The Question

Before:

Summarise this article for me. [pasted]

What came back was a competent summary that told me everything the article contained and nothing about what I should do. Which is exactly what I asked for, so fair enough.

After:

Summarise this article for someone deciding by Friday whether to switch their site off a hosted platform. Only include things that would change that decision.

Same article, unrecognisable output. It dropped two thirds of the content and led with the migration cost paragraph I'd have skimmed past. The lever isn't length or format, it's that a summary has no shape until you say what it's for, and "for me" is not a shape.

What I find interesting about that pair is that the second prompt is barely longer than the first, so the improvement isn't coming from effort or detail in any general sense, it's coming from one specific fact about my situation that the model had no way of guessing and that I'd assumed was implied by the act of asking.

That's the one edit I'd hand somebody who only wants one.

Text Example Two, The Verb

Before:

Can you help me with the about page on my site? [pasted]

Help. The most expensive word in prompting. What came back was a rewritten about page, unrequested, which is not what I wanted, and I'd have got the same result from any model because "help" contains no deliverable.

After:

List the three claims on this about page that a stranger who has never heard of me would not believe. Don't rewrite anything.

Three claims. Two of them fair, one of which I'd been slightly embarrassed about already and had left in anyway. The edit is a verb swap plus a deliverable, and it took four seconds.

I'd note that "don't rewrite anything" carries real weight here and not just as politeness. The default behaviour is to solve, so if you want diagnosis you have to block the solution explicitly.

Text Example Three, Marking The Inferences

This one's less obvious and I use it constantly now.

Before:

Pull the requirements out of this email thread. [thread pasted]

I got a tidy numbered list. Eight items, all plausible, and I only caught the problem because item six mentioned a deadline nobody in the thread had ever typed. It had been reconstructed from context, and it read exactly as solid as the seven items that came from actual sentences.

After, adding one clause:

Pull the requirements out of this email thread. Tag anything you inferred with [inferred] instead of stating it flat.

Three tags came back. Two were fair reconstructions. The third was that deadline again, now visibly marked as a guess, which took it from a trap to a question I could ask someone.

The edit costs you one clause and it turns a confident output into an inspectable one. That's a better trade than almost anything else in this article, and it works on summaries, extractions, translations, and anything where you're going to act on the result rather than just read it.

Image Example One, Why The Object Is There

Now the pictures, where the levers are completely different.

Before:

A wooden desk with a coffee cup and a laptop, morning light.

Catalogue photograph. Everything arranged, nothing touched, the sort of image that appears on a thousand productivity blogs and reads as a set.

After:

A coffee cup pushed to the far edge of a wooden desk to make room for an open laptop, a notebook shoved underneath at an angle, morning light.

The scene got a history. Objects have reasons for being where they are, and specifying the reason drags the model out of the arranged-flat-lay region of its training data and into pictures of rooms people actually use. Over three rerolls of each version, all three of the second batch looked lived in and none of the first did, which isn't a big enough test to publish as fact but was consistent enough that I've kept doing it.

The general form is to describe the relationship between objects, not the list of objects.

Image Example Two, The Light Source You Can't See

Before:

A portrait of an older man in a workshop, natural light.

After:

A portrait of an older man in a workshop, lit by a window out of frame to his left, the far wall in shadow.

Naming a light source that isn't in the picture does something disproportionate. You get directional falloff, a dark side, a plausible room. "Natural light" gets you an even wash from nowhere in particular, which is the visual signature of a render rather than a photograph.

The wider version of this trick is to describe things outside the frame at all. What the subject is looking at, what's casting the shadow, where the sound is coming from. Models handle implication surprisingly well and it costs you five words. The full lighting vocabulary sits in the ai photo prompt guide, which goes much deeper on that slot than I'm going here.

Image Example Three, How Specific The Noun Is

Before:

A car parked outside a house at night.

After:

A beige 1980s estate car with a roof rack parked outside a house at night.

Obvious edit, less obvious result. I expected a more specific car. What I got was a more specific everything, because the era leaked. House got older, the streetlight went sodium orange, the road surface changed, and the whole image relocated itself in time even though I only dated one noun.

That's worth knowing in both directions. A single specific noun will pull the entire scene toward its period and context, which is enormously efficient when you want it and quietly destructive when you don't. If your image keeps coming out with a decade you didn't ask for, look at your nouns before you blame the style terms.

Video Example, Anchoring The First Frame

I'm least confident here, so treat this one as a working note.

Before:

A person walks into a kitchen and makes coffee, morning light.

Two actions and no anchor. What I got back was a figure sort of gliding into a room while the kettle situation resolved itself off in the corner, with the hands going wrong somewhere around the third second.

After:

Opening frame, an empty kitchen in morning light, counter visible, nothing moving. A person enters from the left and stops at the counter. Nothing else in the frame moves. Locked-off camera. Four seconds.

Better. Not good, better. Stating the opening frame gives the model a still to start from, and stating what does not move is apparently as useful as stating what does. The duration in the prompt seemed to help with pacing, though I'd want a lot more attempts before I'd swear to that part.

The Edit Ledger

Everything above plus a few more, collapsed into one table. The last column is whether the effect held when I ran the pair again later, which is the column I wish other people published.

Twelve edits, and the honest summary is that the text ones are reliable and the video ones are hunches. I'd rather show you that split than pretend the confidence is even across the table.

The Edits That Did Nothing

Negative results never get published, which is why everyone keeps rediscovering the same dead ends.

Politeness did nothing measurable. Please and thank you, no effect I could see on quality either way, and on image prompts they're actively wasteful because they occupy space that could carry visual information.

Telling the model to take its time did nothing. Neither did telling it the task was very important to my career, which I tried because people swore by it a couple of years ago.

"Ultra detailed, 8k, masterpiece, award winning" as a block appended to image prompts. I ran a few pairs with and without. I couldn't see a consistent improvement, and on a couple of them the version without was cleaner, though I'd stop short of calling that a result given how few runs it was. What I'm more confident about is that those words spend prompt space, and prompt space on images is finite in a way it isn't on text.

Assigning a photographer role to an image prompt. "You are an award-winning portrait photographer" is a description of a person who is not in the picture, and it did nothing across every pair I ran, which honestly should have been obvious to me from the start given that an image model is matching a caption to a picture rather than following instructions from a hired professional, and yet it's the single most common opening line in every prompt bank I've browsed.

Swapping "purpose" for a different purpose in a template where the role slot was already doing that work. Overlapping slots are wasted slots, and I've cut two from templates I was carrying around for months.

Questions That Come Up

How many examples do I actually need to learn from? Fewer than any bank offers. Eight or ten with the reasoning attached beat four thousand without it, which is my whole objection to the marketplaces.

Can I just use one of the big prompt libraries instead? For visual vocabulary they're genuinely useful, and I say that as someone who complains about them constantly. For text, the finished prompts carry somebody else's context, which is the part that matters. There's more on that in ai prompts.

Does any of this transfer between models? The levers do. The formatting doesn't, entirely, since the vendors recommend different structuring conventions. Structure your prompt the way the model's own documentation suggests and keep the levers the same.

Where do I find more image examples specifically? I've collected the visual ones separately in ai image prompt examples, and the general structure for pictures is in ai image prompt.

Is it worth taking a course for this? Depends what you'd learn instead. I went through what the courses cover and, more to the point, what they skip in prompt engineer course.

Try One Tonight

Take a prompt you sent this week that disappointed you. Not a demo prompt. A real one.

Change exactly one thing from the ledger. Send it again. Then write down what happened in a file somewhere, because the note is the asset and the prompt isn't. Six months from now the model will have moved and the prompt might not work, and "name the decision the reader is facing" still will.

The framework these levers hang off is in chatgpt prompts, if you want the map rather than the examples.