To Prompt Is Human, To Specify Divine
Rethinking how we should interact with Generative AI
Today we treat prompts like throwaway notes scribbled on a napkin. Fire off a quick instruction, read the response, and then type another. But I feel that this mental model — inherited from search engines and old school chat rooms — fundamentally diminishes what prompts could become in the age of generative AI.
Current natural language methods for interacting with generative AI feels like playing twenty questions with an oracle. We start broad (“write a story”), get something unsatisfactory, then narrow down through dialogue: “make it shorter,” “add more conflict,” “change the protagonist.” Chat interfaces reinforce this pattern. Hit Enter to send immediately. Want a line break? Remember to use Shift+Enter — a design choice that signals that prompts should be quick, conversational bursts. (As far as I know, Google’s AI Studio stands alone in treating prompts as persistent artifacts worth editing, not just immediately firing off.)
This ephemeral approach is great for exploratory work, but it doesn’t facilitate developing mature specifications for tasks we perform repeatedly. It also wastes computational resources retreading the same ground. A software engineer wouldn’t rewrite the same function from scratch every time they need it. Yet that’s exactly what we do with prompts.
Consider two scenarios. In the first, a researcher asks an AI to “analyze this dataset and find interesting patterns.” After many rounds of back-and-forth — clarifying what “interesting” means, which statistical methods to use, how to present findings — they get useful results. The next week, facing a similar dataset, they start over with the same vague prompt.
In the second scenario, that same researcher uses the LLM to develop a detailed specification: “Perform exploratory data analysis on [dataset]. Generate correlation matrices for numerical variables, chi-square tests for categorical associations, and identify outliers using IQR method. Present findings as a structured report with sections for summary statistics, key relationships (r > 0.5 or p < 0.01), and anomalies. Include visualizations for top 3 strongest correlations.”
The first approach feels easier initially: unstructured prompts seem like less cognitive overhead. But this trades short-term convenience for long-term inefficiency and computational waste. We avoid the mental work of specifying exactly what we want, then pay that cost repeatedly through extended conversations with the AI. What’s more, the models pay that cost too by processing our redundant explorations over and over again.
But once we’ve refined a prompt through conversation, we possess a mature specification that reliably produces desired outputs. Instead of discarding this knowledge when the chat ends, we should capture and reuse it.
The lesson is that prompts should not be treated as ephemeral communication, but as persistent instruments designed to reliably yield specific results. A well-crafted prompt becomes like a scientific protocol or a software function: something you can invoke repeatedly with confidence in the outcome. Such persistent specifications offer several advantages over ephemeral prompting. They create consistency across multiple generations of the same type of content. They become shareable assets within teams or research groups. And they shift focus from the generation process to the quality of the specification.
When a prompt becomes a specification, the generated output becomes almost secondary. You know you can reproduce it reliably given an appropriate model. The value lies in having precisely articulated your intent, not in any particular instantiation of that intent.
This mirrors how we approach other complex systems. Architects don’t rebuild blueprints for every similar building. Instead they develop pattern libraries and specification templates. (Incidentally, this practice emerged from Christopher Alexander’s pioneering work on patterns in urban design and architecture.) Musicians don’t compose from scratch each time: instead, they build on established forms and structures.
Some may call this approach “prompt engineering,” but what we’re discussing goes beyond the current understanding of that term. Traditional prompt engineering focuses on optimizing individual interactions. True specification-driven AI interaction requires systematic approaches to capturing, refining, and reusing intent across multiple contexts.
Emerging human-centered AI (HCAI) innovations already attack this problem. New interfaces treat prompts as structured, reusable documents rather than ephemeral dialogue. They support prompt refinement over time, version control, and pattern libraries. These tools recognize that sophisticated AI interaction requires moving beyond the chat paradigm. However, even more innovation is needed here.
Software development tools hint at this future. Version control systems, template libraries, and API specifications all recognize that complex instructions deserve careful crafting and reuse. The established areas of formal specification and program verification elevates specification (rather than source code) to the ultimate representation of software. The most advanced AI interaction tools should embrace similar principles — and increasingly do.
So, to prompt is human, to specify divine. The difference lies not in the initial exploration — that conversational dance will always have value — but in recognizing when we’ve discovered something worth preserving. In other words, the most sophisticated AI users won’t be those who craft the cleverest one-off prompts, but those who build comprehensive specification libraries that reliably produce desired outcomes.
