Why Prompt Templates Don’t Work

Written by Irakli Eliava

We have told you, more than once, that prompt engineering is not the thing to focus on. Then we looked at what 1,678 real briefs sent through our own tool were actually for, and 224 of them, 13.4% of everything, one in eight, were people writing prompts. Agent instructions. System prompts. Prompt libraries for other AI tools. The second largest use case in the entire dataset, behind only general content writing. Either we were wrong to say that, or the people using our tool were wrong to do it. We do not think either is true, and the data itself explains why.

What We Told You, and What 224 People Did Anyway

Three posts on this blog argue that prompt engineering is overrated: that structured prompt formats and template libraries are a distraction from the work that actually matters. We stand by that argument. We are not retracting it. What we had not done, until we read our own usage data, was check it against what our users actually do here. Grouped across every brief sent through Briefing Fox in six months: “AI and prompt engineering” is 224 briefs, ahead of business strategy, ahead of every academic use case, ahead of all image and video generation combined. Inside the developer category alone it shows up again at 10.5% of everything developers ask for. That is not a fringe behaviour we can wave off. It is the second most common thing this product is used for, running directly against what three of our own articles say people should be doing instead.

The Category That Fails Three Times as Often

Here is the number that resolves the contradiction rather than deepening it.

Category	Briefs	Abandoned	Rate
Text and Logic	1,065	107	10.0%
Developer and Coding	384	66	17.2%
Image Generation	94	16	17.0%
Agentic AI and Workflows	72	24	33.3%

Agentic AI and Workflows, the category that covers writing prompts and instructions for other AI systems, abandons at three times the rate of ordinary text requests. One brief in three never makes it through. It is the smallest major category in the dataset and by a wide margin the worst performing. That is not evidence that prompt engineering is a bad idea. It is evidence that specifying an agent is genuinely harder than specifying a document, and that most people attempting it have no method for doing so. They reach for a template because a template looks like a method. It is not one.

Templates and Briefs Are Not the Same Move

This is the distinction the three earlier posts on this blog were gesturing at without quite landing on it: prompt templates fail for the same reason every borrowed framework fails. They were built for someone else’s agent, someone else’s task, someone else’s edge cases. Filling in the blanks does not add the one thing a working agent prompt actually needs, which is a clear, specific account of what this particular agent is for and what it is allowed to assume. A brief is different from a template in exactly one respect that matters: a brief starts from your situation and asks what an agent built for it would need to know. A template starts from a stranger’s situation and asks you to adapt. Adapting someone else’s answer is slower and worse than answering the question yourself, and it is the reason a category built entirely around specifying AI behaviour loses a third of the people who try it.

What a Briefed Agent Prompt Has That a Template Never Will

An agent instruction copied from a library carries someone else’s assumptions about scope, tone, error handling, and what “done” looks like. None of those assumptions are stated, because the template’s author never met your task. When the agent behaves oddly, there is nothing in the prompt to diagnose, because nothing in it was actually decided by anyone thinking about your case. A briefed prompt states those things because someone was made to answer them before writing a word. What should this agent refuse to do. What does it need to know that is not obvious from the task name alone. What happens when it is uncertain. Where does it stop and hand back to a person. Those are not stylistic additions. They are the actual content of a working agent instruction, and a template has no way to contain them because it was written before your agent existed.

Before You Write Your Next Agent Prompt

Skip the library. Before you write the instruction, answer four questions in plain language: what is this agent actually for, what should it never do, what does it need to know that you have not written down anywhere else, and what should happen when it hits something it cannot handle. That is a brief, not a template, and it is the difference between the 66.7% of agent-prompt attempts that finish and the third that do not. Try Briefing Fox free at briefingfox.com