When teams build AI prompts together, who ends up controlling which part — and does that blur the line between design and engineering?

This explores how teams prototyping with generative AI split a prompt into distinct parts, and why that split matters. The most direct evidence comes from a study of 39 practitioners across 13 design sessions. Teams kept arriving at the same four-part anatomy: **input context** (the material the model works on), **system instructions** (what it should do and how it should behave), **output constraints** (format, length, tone), and **few-shot examples** (sample input-output pairs) How do teams decompose prompts when collaborating on generative AI?. The more surprising finding concerns who touches what. Designers, engineers and domain experts each gravitated toward different components, so the prompt became a shared workspace that blurs the usual line between design and engineering work.

The four-part split is more than a convenient filing system, because other research shows the parts respond differently when you change them. Prompt compression work found you can shrink prompts up to 20x with little accuracy loss, but only if you protect instructions and questions and cut much harder into the examples Can prompts be compressed without losing model performance?. In practice, the instructions carry most of the weight, and the examples are often padding. A team that has already separated its components knows where it can trim and where it can't.

A second reason to keep the components apart is that problems become easier to diagnose. One framework scores prompt quality on six dimensions, including clarity of communication, cognitive load, instruction design and logic, and finds that fixing one dimension often improves others Can we measure prompt quality independent of model outputs?. Separate components give a team a natural place to apply those checks: a vague output constraint is a different fix from a confusing instruction. There is a caution, though. Even well-separated components are sensitive to exact wording. Prompts that mean the same thing can produce noticeably different output, because models respond to how common a phrasing was in their training data rather than to its meaning Why do semantically identical prompts produce different LLM outputs?. A modular prompt still needs each module tested, and which techniques help also depends on the model tier Do prompt techniques work the same across all LLM tiers?.

The less obvious angle is what decomposition can't fix. One analysis argues that a prompt packs the request, the background context and the model's role into a single fixed frame that the model can't renegotiate the way a human conversation partner would How do prompts reshape the role of context in AI conversation?. Splitting that frame into parts makes the packing visible and editable for the team, but the model still receives it as one block. Iteration has a further side effect. Each round of refinement steers the output toward what the team already expected to see How much does the user shape what a model generates?. So a well-organized prompt can make it easier for a team to confirm its own assumptions. Some designers propose showing a preview of the likely output before a prompt is submitted, which would surface mismatches earlier Can previewing AI output before submission reduce prompt iteration?.

The corpus has limited material here. Only one study looks directly at how teams decompose prompts during prototyping. The rest addresses the same territory from compression, evaluation and interaction design, so treat the four-part model as one well-observed pattern rather than a settled standard.