Minimum Viable Prompt: Your Cure for AI Overwhelm

How embracing the "good enough" helps you dive in.

Happy Thursday, web wizards!

You might have noticed a theme in my recent Thursday posts: demystifying AI.

This is intentional.

It’s my counter-movement to crap like this:

• "ChatGPT is a miracle, but 99% of you are using it WRONG! ❌"

• "Everyone is using the latest [AI TOOL], and you're FALLING BEHIND! 🧗"

• "If you don't replace your physical body with an AI-powered robot, you will LITERALLY DIE! ☠️"

To be fair, that last one is true in the long run.

All of this noise helps “AI Gurus” get rich at the expense of making generative AI appear more inaccessible than it really is.

I call bullshit.

So today, I propose a simple mindset shift to reduce AI anxiety and bring back the magic.

The case for the Minimum Viable Prompt

Here’s something I’ve come to believe.

What separates people who treat generative AI as unapproachable and those who find it intuitive isn’t their level of technical skills, prior knowledge of machine learning concepts, or exclusive access to secret hacks¹.

No, it’s their willingness to try shit out and learn by doing.

I’ve already advocated for a “start simple” approach when it comes to text-to-image models over a year ago.

Today, I’m introducing a new term for this test-and-see way of working with generative AI: “Minimum Viable Prompt.”

(Because we don’t have enough “MVP” acronyms competing for our brain space².)

What is the “Minimum Viable Prompt”?

MVP is just a prompt that covers your bare minimum requirements.

Shocking twist, I know!

Now, your bare minimum requirements will vary, so MVP doesn’t always mean “short and sweet.”

For instance, here’s what a Minimum Viable Prompt might look like:

Dog

But it could also look like this:

Source: One Useful Thing

Ethan Mollick

’s “Student Coach” prompt above is the result of experimentation followed by lots of tweaking. It presumably does exactly what Ethan needs it to do, no more, no less.

But if all you’re looking for is a random picture of a cute dog to insert into a blog post about animals, “dog” is plenty.

For most beginners, MVP will be closer to the latter. Over time, your MVP is likely to scale with your level of experience.

MVP forces you into an exploratory mindset.

Instead of asking:

“What is the best possible prompt for this purpose and where do I find it?”

You ask:

“What do I need and how do I request it as simply as I can?”

The idea of a Minimum Viable Prompt applies to most types of AI interactions, whether it’s with a chatbot like ChatGPT or an image model like Midjourney.

Why use a Minimum Viable Prompt?

Taking the MVP approach does several good things:

1. It gets you going

If you start with a “good enough” mindset, you get to avoid analysis paralysis.

Instead of agonizing over which perfect prompt to pick, you simply dive in and see how things work. Then, after you’ve gotten the first result, you can adjust and improve your prompt.

Don’t underestimate the power of inertia!

2. It lets you gauge the model’s capabilities

This is relevant if you’re testing a new AI model and want to compare it to another one you’re familiar with.

A long, detailed prompt might “mask” the model’s default behavior, making it trickier to figure out whether it’s your prompt or the model itself doing the heavy lifting (see the worked example in the next section).

A more basic prompt lets you get a feel for how the model acts before building on it.

3. It helps you learn by doing

Starting small lets you test the waters and see what works and what doesn’t. You get to develop your skillset as you evolve your prompts iteratively.

This is especially true if you combine your minimum viable prompt with something like the “ask me questions” method. The model’s response will showcase what it’s capable of and walk you through subsequent interactions.

4. It builds a foundation for advanced methods

You can think of MVP as a tutorial in a video game.

If you’re a first-time gamer, playing through the first few levels will be far more beneficial than having an experienced player hand you their 80-level Warrior Orc, equipped with all sorts of gear you’ve never seen before³.

This is the equivalent of learning to fish vs. getting a fish handed to you.

Learning the ropes puts you in a much better position to appreciate the more elaborate prompting techniques and understand their usefulness.

Worked example: The oak photo

Say you need a photo of a tall oak for your magazine, I Like Big Trees (And I Cannot Lie)⁴.

You know that Midjourney is great at photographic images, but you’ve never used it before. After reading multiple posts by Midjourney experts and looking through several prompt databases, you eventually land on this off-the-shelf prompt you saw someone else use:

“tall oak, 64K, HDR, high resolution, highly detailed, extreme detail, masterpiece, professional camera, award-winning photography”

You copy-paste that into Midjourney…and sure enough, there’s your tall oak:

That is one pretty oak photo!

But here’s the thing (or four):

1. You don’t know how much of the resulting image is due to Midjourney’s default aesthetic and how much is driven by your long prompt.

2. You don’t know how many of the modifiers are necessary for the image to look like that.

3. You don’t know which modifier has what kind of effect.

4. You’re now likely to continue using the above prompt as your starting point for future Midjourney photos, blindly copy-pasting the entire string as if it’s set in stone.

But what if you went with the Minimum Viable Prompt instead?

Remember, in our hypothetical case, all you’re after is a photo of a tall oak.

So that’s where you start:

That’s…a very similar shot with a 4X shorter prompt:

MVP vs. Copy-Paste

You now know exactly the kind of default image you can expect from Midjourney without additional descriptors.

From here on out, you get to build on your prompt according to your needs.

If you decide that you’d like a view from the side instead of the bottom-up shot, your MVP becomes:

“photo of a tall oak, side view”

Want a night setting? Here’s your new MVP:

“photo of a tall oak at night, side view”

You get the picture.

The best part is that you can take your clean MVP into another text-to-image model⁵ and learn if it behaves differently.

Here’s vanilla SDXL:

Curious.

Looks like SDXL tends to use the side view by default, so you don’t even need the extra “side view” modifier.

If you ask me, it’s more fun and liberating to take the pressure off with this less-is-more approach.

Over to you…

What’s your take on AI prompting?

Is the Minimum Viable Prompt concept already painfully obvious and natural to you?

If you have other ideas to help smooth the learning curve for beginners, I’m all ears.

Leave a comment or shoot me an email at whytryai@substack.com.

1. I know this because I’m a perfectly average tech user with no access to secret Illuminati hacks who still couldn’t tell you exactly how a lot of this AI stuff works.

2. Take a seat, “Most Valuable Player” and “Minimum Viable Product.”

3. Ethan Mollick is the 80-level Warrior Orc player in this analogy, I guess.

4. Alternative: Baby Got Bark

5. We may have more text-to-image models to pick from, but they’re all getting increasingly better at responding to natural language.