Prompts are specs. Write them like one.

An Instagram reel this week by Cooper Simson ↗, an AI content creator who goes by @cooper.simson, declared prompt engineering "the biggest waste of time in AI." The argument: models got smart enough that long, detailed prompts now get in their way. His fix is a four-part shape he calls the pyramid, gated behind a "comment PYRAMID for the template" hook. Engagement bait aside, we watched it, and the annoying thing is that he is mostly right.

The pyramid, decoded

Strip the animation and the framework is four layers:

• Goal. Describe the finished thing, not the steps to make it.
• Research. Let the model survey the options and pick the best way.
• Interview. Make it ask you questions first: who is this for, what does good feel like, show me an example.
• Example. Give it one real sample of what good looks like.

Then his last beat, which most people will skip: share the resulting prompt with your entire team.

Here is the part that made us laugh. That is not a prompting technique. That is a scope document. We run the same four moves on every client project: figure out what done looks like, survey the approaches, ask the client questions until the vague parts are specific, and anchor everything to a real example. He did not invent a prompting framework. He reinvented discovery.

Why it works, in spec terms

The old style of prompt engineering was ceremony. "You are a world-class expert. Take a deep breath. Think step by step." That is ritual, and models did outgrow it. But length was never the actual problem. A prompt can be long for two reasons: information or noise. Cutting noise always helps. Cutting information only helps when the model already knows it.

The pyramid works because each layer replaces noise with the specific information a spec carries:

• Goal is your acceptance criteria. "The finished thing" is what a QA team would test against.
• Research is letting the model choose the method, the way you would let a senior engineer pick the approach instead of dictating it.
• Interview is the discovery call. The questions in the reel, "who is it for" and "show me one," are literally the questions we ask on kickoff calls.
• Example is a reference implementation. One real sample beats three paragraphs of description, the same way one screenshot beats a page of UI notes.

So the reel's real claim should have been: prompt ceremony is dead. Write prompts like specs and the details still matter.

Where it breaks

Two honest limits, because a guide that skips them is marketing.

The interview step assumes a human in the loop. In production, there is no human to interview: a prompt inside a pipeline needs a deterministic output, which means you are back to explicit contracts anyway. Output schemas, field names, error cases. When we ship AI features, like the on-device insights in CashTaq, the prompt is a deliverable with tests around it, not a conversation. The pyramid is for the messy front half of work, the planning and drafting and deciding. The back half still wants a contract.

And "prompt engineering was a waste of time" is the kind of sentence designed to be clipped. The discipline was never about magic words. It was about knowing what the model needs to know. That part did not go anywhere.

The template

Steal this. It is the pyramid rewritten the way we would write it internally:

GOAL      what done looks like, one sentence a reviewer could check
RESEARCH  "survey the approaches and tell me which you would pick and why"
INTERVIEW "before you answer, ask me the questions you need answered"
EXAMPLE   paste one real sample of good output and say "match this"

The one-line version for lazy days: "Here is the goal. Ask me what you need, then match this example."

The takeaway

The best prompt advice moves your thinking from how to what. The pyramid does that, which is why it beats the old ceremony. But do not throw away the habit underneath: be specific about done, honest about constraints, and generous with examples. That habit has a name already. Around here we call it a spec, and we have been charging clients for it for years.

Also, ship your prompts like code. His last point deserves more than a footnote: a prompt the whole team shares, reviews, and improves is an asset. A prompt one person keeps in their head is a liability.

← the journal