Every few months there's a new debate about whether LLMs have plateaued. Better benchmarks, bigger context windows, cheaper inference. And yet most people using ChatGPT, Claude, or Gemini day to day are still getting mediocre output.
The uncomfortable truth: for the vast majority of real-world use cases, the model stopped being the bottleneck a while ago. The prompt is.
The variance is bigger than you think
I've run the same task across GPT-5.x, Claude, and a couple of open-weight models, and the gap between a lazy one-line prompt and a properly structured one is often larger than the gap between the models themselves.
Type "write me a marketing email" and you get something forgettable, regardless of which model answers. Give it a role, real context, a defined objective, and explicit constraints, and the exact same model performs like it was actually briefed.
This isn't news to anyone who's spent real time with these tools. But it's still not reflected in how most people actually use them.
Why structure gets skipped
Frameworks like CO-STAR (Context, Objective, Style, Tone, Audience, Response) and RISEN (Role, Instructions, Steps, End goal, Narrowing) exist because a good prompt has real, repeatable parts. The problem is holding all of that in your head every single time you open a chat window.
So people default to the path of least resistance: a single sentence, sent cold. Then the output is underwhelming, and the conclusion becomes "AI isn't that good," when the actual issue was never given enough to work with.
A concrete example
Here's the same task, unstructured vs structured:
Unstructured:
Fix this code.
Structured:
Act as a Senior [LANGUAGE] Engineer reviewing a pull request.
Context: This function is supposed to [EXPECTED_BEHAVIOR] but instead [ACTUAL_BEHAVIOR].
Objective: Identify the root cause and propose a fix.
Format: Explanation first, then the corrected code block.
Constraints: Do not rewrite unrelated code. Flag any edge cases you notice.
Same model. Same underlying capability. Completely different quality of output, because the second version actually tells the model what "good" looks like.
The real fix isn't "learn to prompt better"
Telling people to just get better at prompting doesn't scale, the same way telling people to "just write better emails" doesn't fix inbox chaos. What actually moves the needle is making structure the default instead of something you have to remember and reconstruct every time.
That's the gap most prompt tooling is trying to close right now: not smarter models, but better scaffolding for how people talk to the models they already have. A saved template, a builder that walks through the missing pieces, a shared library so a team isn't rebuilding the same prompt from scratch every week.
If you're curious what that looks like in practice, I put together a structured prompt writer and template library built around CO-STAR, RISEN, and CRAFT, covering 29 categories from coding to marketing to HR, all reusable across ChatGPT, Claude, and Gemini.
Where this is heading
As more real work runs through these models, the ability to brief them clearly is going to matter the same way clear technical writing already does. It won't be a nice-to-have skill for power users. It'll be assumed, the same way basic Git literacy is assumed on most teams today.
The models will keep getting better regardless of what any of us do. The more interesting problem, and the one actually worth solving right now, is closing the gap between what these models can do and what most people are actually getting out of them.
What's your experience been? Curious if others have noticed the same variance between prompt quality and model choice.
Top comments (0)