Even a Mediocre Prompt Can Build Pretty Good Software Today

Here’s the exact prompt I typed into Replit the other day

“ok i have an idea where we built a new speaker card page, where if you upload a headshot, we build you a 1080×1080 card on a background you pick from (one is attached) and show your headshot and ‘I’ll Be at SaaStr AI 2026’ or can tweak text”

That’s it. One run-on sentence. No capitalization. No structure. I attached a background image and hit send.

Five minutes later, the AI agent had built the full page — headshot upload, circular crop with a glowing border, three background options, editable text, 1080×1080 PNG export — wired it into the site router, added a footer link, verified the build was clean, and deployed it to production.

The app is live right now at saastrannual.com/attendee-card. Thousands of attendees will use it to generate LinkedIn cards for SaaStr Annual. It’s a real product feature on a real website, built from a prompt I wouldn’t have given a passing grade in a middle school writing class.

I think we’re past prompt engineering. And I don’t think most people have caught up to that yet.

Yes, the agent and I spend the next 20 minutes fixing bugs and getting it into production. But a mediocre prompt by me didn’t stop us. Not one bit.

The Prompt Engineering Era Is Already Over

For the last two years, there’s been an entire cottage industry built around prompt engineering. Courses. Certifications. Job titles. LinkedIn influencers posting about “the perfect prompt framework” with acronyms like C.R.E.A.T.E. or R.I.S.E.

That advice wasn’t wrong at the time. Early LLMs were brittle. You had to coax them into useful output. The difference between a bad prompt and a good prompt was often the difference between garbage and something usable.

That gap has collapsed.

The models got dramatically better at inferring intent from messy, casual, incomplete input. You don’t have to structure your prompt perfectly. You don’t have to specify every constraint upfront. You don’t have to use magic words. You just have to say roughly what you want, in whatever way you’d naturally say it, and the AI fills in the rest.

My prompt didn’t specify Canvas 2D API rendering. It didn’t mention CORS proxy endpoints. It didn’t describe the headshot cropping algorithm. It didn’t say “wire it into the Next.js router and add a footer link.” I said “build a card page where you upload a headshot” and the agent figured out every implementation detail on its own.

What Actually Matters Now

If prompt engineering is dead, what replaced it?

**Knowing what you want to build.**The bottleneck was never typing the right words in the right order. It was having a clear product idea. I knew exactly what this card generator should do because I’ve been thinking about attendee engagement for a decade. The AI can’t give you that.**Recognizing when the output is wrong.**The first version of my card generator had a broken download — the SaaStr logo wouldn’t render in the exported PNG because of a cross-origin browser security issue. I didn’t need to know what CORS was. I just needed to look at the output and say “the logo isn’t showing up in the downloaded image.” The AI diagnosed it and fixed it.**Taste.**I chose the backgrounds. I decided how big the headshot circle should be. I picked the cyan glow effect. I decided the text should default to “I’ll Be at SaaStr AI 2026” with a 50-character limit. These are product decisions. Design decisions. The AI executes them well, but it doesn’t make them for you.**Willingness to iterate.**The first version had 3 backgrounds. We later added 2 more. The first headshot crop centered on the middle of the image, which cut off people’s faces. I said “faces are getting cut off” and the AI shifted the anchor point to the top of the image. Two rounds of feedback. Not two sprints. Two sentences.

None of this requires prompt engineering. It requires the same skills that made someone a good product manager or founder in 2015. The difference is you don’t need a team of engineers between your idea and the shipped product anymore.

This Further Changes Who Can Build Software

The uncomfortable implication: if you don’t need to be a prompt engineer, and you don’t need to be a software engineer, then the number of people who can ship production software just went from a few million to a few hundred million.

We’re already seeing this at SaaStr. We run the entire company with 3 humans and 20+ AI agents. I’m not an engineer. I took CS classes in college, but I haven’t written production code professionally in over 15 years. In the last year, I’ve shipped 10+ production apps on Replit — including internal tools, public-facing features, and a startup simulation game that hit 47,000+ lines of code.

I’m not prompt engineering my way through any of this. I’m describing what I want in plain English and course-correcting when the output misses. That’s it.

The skill that matters in 2026 isn’t writing good prompts. It’s having good ideas and good judgment about what to ship.

The prompts can be mediocre. The taste can’t be.

The Real Test

Here’s how you know prompt engineering is over: go build something. Don’t watch a tutorial on prompt structure. Don’t use a framework. Don’t optimize your wording. Just open Replit or Cursor or Claude and describe what you want like you’re texting a friend.

If the output isn’t right, say what’s wrong. Not in technical terms. Just “the logo isn’t showing up” or “the text is too small” or “it should look more like this.”

You will be surprised how far you get. And the gap between what you can build with a “perfect” prompt versus a casual one-sentence description? It’s almost nothing now.

We spent two years treating prompts like incantations — as if the exact arrangement of words was the secret. It wasn’t. The models just needed to get good enough to understand what we actually meant.

They’re good enough now.