How do teams decompose prompts when collaborating on generative AI?

Observing how 39 practitioners (UX researchers/designers, software developers, product managers) authored prompts across 13 collaborative design sessions, the study identifies four recurring sub-components of a prompt: "(1) the input context, (2) the system instructions, (3) the output constraints, and (4) few-shot examples." Input context supplies "the necessary background or situational context" the end-user would provide; system instructions are "commands or requests detailing what the model is expected to do"; output constraints are "guidelines or boundaries for the generated response," covering both stylistic choices like format and length and higher-level guardrails against harmful content; few-shot examples give the model "illustrative content that implicitly indicated the desired nature of the output."

The study also maps roles onto these components. Designers, with expertise "crafting interfaces and designing input/output flows," gravitated toward treating the prompt as a metaphorical "template." Some design participants, by contrast, "encountered challenges in understanding the concept of fewshot examples" which the authors describe as "fundamentally rooted in engineering principles." Because the four sub-components are interconnected — refining one tends to require revisiting the others — the authors find that prototyping work "blurred lines between roles and design representations," obscuring "individual agency and ownership" and challenging the conventional, modular division of design labor. They invoke Subramonyam et al.'s concept of "leaky abstractions," which allows some overlap across role boundaries while preserving distinct artifacts per role, but argue that in generative AI prototyping the separations become "even more indistinct," with shared artifacts further dissolving role lines. The authors leave as an open question how roles, design representations, and prototyping methodologies should adapt to this.

This four-part decomposition of the prompt artifact runs parallel to How should generative UI be organized as a design space?, which proposes a four-axis taxonomy for the generative UI design space as a whole (audience, constancy, control, timeframe); here the four parts decompose the prompt itself as the unit teams actually build and hand off between roles, rather than the broader interaction it produces. It also offers one account of the gap described in Do generative UI tools actually implement their stated design rationales?: if output constraints and few-shot examples are the components non-engineering roles find hardest to specify, that is a plausible source of the roughly one-in-four design rationales that benchmark finds unimplemented, though this study does not measure implementation rates itself. The underlying content-centric practice this anatomy supports is described in the sibling note Does generative AI prototyping focus on model architecture or content curation?.

The four-component anatomy and the role-blurring it produces come from a single domain (marketing content creation) and a single tool (AI Studio), observed across 13 sessions; the authors explicitly flag the role question as unresolved rather than answered, and note separately that prompt-level transparency issues "may be hidden at the prompt instruction level" from end-users without demonstrating how often this occurs. The excerpt supports the four-part anatomy as a descriptive framework for this study's setting; whether the same decomposition or the same role-blurring recurs with other prompting tools, domains, or larger teams is not established.