The prompt was never the valuable part. The thinking behind it was.
I’ve been building AI workflows for over two years. In that time I’ve watched an entire industry spring up around selling prompt templates, and I think most of it is built on a shaky foundation.
Wharton’s Generative AI Labs put this to the test. They ran “You are a world-class expert” persona prompts across six models, tens of thousands of times. Expert personas had no significant impact on accuracy.1 Mismatched experts sometimes made things worse. The tricks that template sellers charge for, the persona assignments and specific phrasing and structure, didn’t survive contact with rigorous testing.
That tracks with what I’ve seen firsthand.
In 2024, when I started working seriously with AI, I believed the prompt was the thing. Get the wording right, the structure tight, the instructions precise, and you’d get consistent results. Sometimes that worked. But just as often, the same prompt gave me something completely different on the next run. Inaccurate, unfocused, or disconnected from what I was trying to do. What I figured out over time was that the prompt only did heavy lifting when it had real context behind it. Reference documents, research, brand guidelines, project background. Without those, even a good prompt missed more than it hit.
What changed
By 2026, the prompt became the smallest part of how I work. The context around it does almost everything now. When I set up a workflow, the prompt is a plain explanation of what I’m trying to accomplish, what needs to be included, and what I don’t want. No personas. No magic phrasing. The real work lives in the workspace: project instructions, reference materials, accumulated context from everything I’ve already done in that project.
The broader field caught up to this too. Researchers started calling it “context engineering” instead of “prompt engineering,” because the data keeps showing that what you feed the model matters more than how you ask. Sclar et al. found that formatting changes alone can swing accuracy by up to 76 points.2 Another team documented a “butterfly effect” where tiny edits that don’t change meaning at all can flip the output.3 If a prompt is that sensitive to surface-level changes, packaging one up and selling it as a finished product doesn’t make a lot of sense.
How I actually build things
My process has four parts, and the prompt barely registers as one of them.
- I start by thinking through the idea. Not asking AI to produce anything, just going back and forth in conversation to get clear on what I’m actually making. This is where the real work happens, and it’s the part that every template skips.
- Then I build a plan. I pull in research, brand guidelines, format requirements, whatever context is relevant to what I’m creating. If I’m building a social media post, the plan includes the data the post draws from, the voice it needs to match, and the direction it’s heading. A plan built on evidence will outperform a clever prompt running on nothing, every single time.
- From there I create the output. With enough context loaded, the first pass is usually close. Not done, but close. The gap between starting from context and starting from a template is that one sounds like it came from your business and the other sounds like it could’ve come from anyone’s.
- Then I humanize it. I go through and fix what doesn’t sound like me, adjust the design, catch the spots where the writing feels generated. AI handles most of the distance, but the last stretch is where your voice and your judgment finish the work. I don’t think AI should run a full workflow untouched. But getting to a point where you can build on what it produces, that’s worth the effort.
Why templates break down
When you buy a prompt template, you’re buying someone else’s context stripped out and replaced with fill-in-the-blank fields. Your brand, your audience, your constraints, your standards, none of that comes with it.
The data backs this up at scale too. McKinsey’s 2025 State of AI survey found that organizations actually succeeding with AI are 2.8 times more likely to have fundamentally redesigned their workflows, and only about 20% of other companies have done that at all.4 The gap isn’t about access to better prompts. It’s that most people and most organizations are bolting AI onto what they already do instead of rethinking how the work gets done.
There’s a judgment problem on top of that. A Harvard and BCG study tracked 758 consultants using AI and found that on tasks inside AI’s strengths, output quality jumped about 40%. But on tasks outside that range, people who used AI were 19% less likely to get the right answer than people who worked without it.5 A template can’t tell you which side of that line your task falls on. Only your own understanding of the work can.
And there’s a sameness problem. Doshi and Hauser found in a 2024 Science Advances study that AI boosts individual creativity but makes outputs 10.7% more similar to each other.6 The more people leaning on the same shared templates, the more everyone’s work starts converging. The thing that makes your output recognizable gets averaged away.
I’ll be fair though. Templates aren’t useless as a starting point. If you’ve never worked with AI before, a template can show you what a prompt looks like and get you moving. The problem is staying there. Treating the template as the destination instead of the on-ramp means you never build the context layer that makes AI actually useful for your specific work.
If you build your own process instead, going back and forth with AI, pulling in your own research, planning around your own context, the output carries your thinking in it, shaped by your constraints and finished by your judgment. A prompt pack can’t give you that, because the valuable part was never the prompt. It was the thinking behind it.
Sources
Basil, Shapiro, Shapiro, Mollick, Mollick & Meincke, Prompting Science Report 4: Playing Pretend: Expert Personas Don’t Improve Factual Accuracy, Wharton Generative AI Labs (December 2025). gail.wharton.upenn.edu ↩︎
Sclar, Choi, Tsvetkov & Suhr, Quantifying Language Models’ Sensitivity to Spurious Features in Prompt Design, ICLR 2024. The 76-point gap was measured on one model (LLaMA-2-13B). arxiv.org/abs/2310.11324 ↩︎
Salinas & Morstatter, The Butterfly Effect of Altering Prompts: How Small Changes and Jailbreaks Affect Large Language Model Performance, Findings of ACL 2024. aclanthology.org ↩︎
McKinsey & Company, The State of AI in 2025: Agents, Innovation, and Transformation (November 2025). High performers: 55% had fundamentally redesigned workflows, against 20% of other respondents. mckinsey.com ↩︎
Dell’Acqua et al., Navigating the Jagged Technological Frontier, Organization Science (2026); working paper SSRN 4573321 (2023). The study reports the gap as 19 percentage points. Summary from Harvard Business School · Working paper ↩︎
Doshi & Hauser, Generative AI Enhances Individual Creativity but Reduces the Collective Diversity of Novel Content, Science Advances 10(28) (July 2024). doi.org/10.1126/sciadv.adn5290 ↩︎