#25 | Is prompt engineering dead?

Hey there,

Remember when prompt engineering felt like solving a Rubik’s Cube? I do.

A while ago, I was neck-deep in prompt engineering.

I’d built a tidy little system: outline → content draft → tone refinement. It looked solid—structured, repeatable, and kind of elegant.

But in practice? It was still fragile. One rule change and the output lost its shape—tone, structure, everything.

At one point, I spent 40 minutes tweaking a single instruction just to stop the AI from starting a sentence with ‘Ever.’ AI loves to ignore rules, so human.

Eventually, I gave up and typed this into Claude 3.7:

“Can you clean this up and make it feel more persuasive, honest, and slightly annoyed with the usual productivity advice?”

That’s all. No roles. No tone samples. No stacked rules. Just a clear prompt in plain language.

And it worked. Not perfectly. But better. Quicker. Easier to revise.

I then thought:

Maybe prompt engineering isn’t dead. But maybe it’s not the mega skill we (I) thought it was.

Let’s find out.

The AI Learning Guy newsletter 🤖 🧠💡

AI learning hacks and mega prompts delivered to your inbox.

Back when prompts ruled everything

If you were using AI between 2022 and early 2024, you probably remember:

Prompt templates. Notion libraries. Mega-prompts built like flowcharts. People stacking logic just to get a decent draft out of GPT-3.5 or Claude 1.0.

For a while, it worked.

We had to force the outputs to stay on track because the models couldn’t carry context, remember tone, or adhere to constraints.

So, we built systems around the gaps.

Prompt engineering was a workaround. And at the time, it made sense.

But that’s not where we are now.

2025 AI doesn’t need that much help

Let’s call this what it is: a shift. The tools changed quietly—and fundamentally.

Here’s what’s different in 2025:

Claude 3.7 holds tone and context across a session—no constant reminders.
ChatGPT 4.5 adjusts mid-reply if the tone drifts.
Gemini 2.0 Flash prefers shorter prompts. Leave space, and it fills the gaps.
Midjourney v6 lets you prompt in plain language and still get near-usable results.

We’re not prompting brittle prototypes anymore.

We’re briefing systems that understand structure, tone, and follow-up.

Often, the more we try to engineer, the worse the output gets.

When prompts still work—and when they break

Prompt design still has its place. But it’s more selective now.

It still works well for:

Tasks with layered constraints (e.g., format + tone + voice)
Visual prompting that needs layout or style guidance
Structured documents (e.g., outlines, tables, workflows)
Reasoning or multi-step logic tasks
Repetitive tasks using the same instructions

It tends to break when:

You rely on chains that collapse with one change
You over-instruct and under-iterate
You spend hours refining a prompt instead of reviewing the output

Here’s a real example:

I used a 3-part content generator for a project:

Insert voice persona
Apply section structure via markdown
Inject tone samples to match brand style

It worked—until it didn’t. Any time I adjusted the topic, it slipped. The tone flattened, or the structure drifted.

Now I do this instead:

“Write this like someone who’s tired of fluffy advice but still trying to be helpful. Direct, clear, no cheerleading.”

It’s faster. It holds together. And it’s easier to iterate on. (Remember to always iterate!)

What matters now: clarity over complexity

Prompt engineering used to be about controlling the machine.

Now, it’s about communicating with it.

We’re not stacking instructions but rather giving direction.

Here’s the (mental) filter I use every day:

The Prompt Simplifier Test

Can I explain this in two sentences to a colleague in Slack?
Will the model understand my intent based on tone, task, and phrasing?
Is it faster to prompt simply and edit than to build a multi-step chain?

If yes, I skip the stack and write the clear version.

One follow-up usually beats ten setup lines.

Try this for yourself.

Take one of your old “reliable” prompts—the one you’ve rewritten a dozen times.

Now, simplify it.

Original:

“You are a content strategist. Structure this article in four parts with markdown. Include clear CTAs. Maintain a professional, informed tone. Use short sentences. Avoid exclamation points. Output should follow SEO best practices…”

New:

“Structure this in four sections with one CTA at the end of each. Keep the tone helpful, clear, and grounded.”

Run both in Claude or GPT-4.5. See which one gives you a better starting point. Of course, you can use AI to help you simplify it, but be aware of unwanted changes or omissions.

So… is prompt engineering still worth it?

If we talk about scripting instructions.

When the task is complex, structure matters, and context alone won’t carry the output.

But most of the time?

We need clarity. We need tone. And we need to know how to guide the model—not force it.

If your prompt takes longer to engineer than the output takes to fix?

That’s your signal.

Catch you in the next one,

Mark
The AI Learning Guy
👋⚡😎

P.S. Try the two-line version of your longest prompt. Let me know what surprised you.

The AI Learning Guy newsletter 🤖 🧠💡

AI learning hacks and mega prompts delivered to your inbox.

Learning Ressources

Note: No single website has all the answers. This list offers a starting point for those who want to roll up their sleeves or simply satisfy some AI curiosity. Links: No affiliate links included.

Stay connected – I’m here to help!

Have questions? Need support?

Have topic ideas? Want to share your success story?

I’d love to hear from you. Your feedback will also help me improve the newsletter.

Feel free to reach out anytime.

The AI Learning Guy
Contact: Leave a message.