IBM watsonx Prompt Lab is where teams experiment with generative AI models by writing prompts, tweaking parameters, and evaluating outputs before deployment. But right now, the workspace only supports one prompt, one model, and one pane at a time.
If a user wants to compare two phrasings or two models, they must run one variation, copy the output elsewhere, overwrite their prompt, and run the next. This creates continuous friction during prompt engineering.
This friction point was identified through hands-on workflow analysis of prompt engineering environments. Generative AI iteration is not a linear task, it is inherently comparative. Forcing users to overwrite active prompts or switch context to compare two outputs is a clear usability bottleneck for technical teams.
Before deciding on a side-by-side canvas, I evaluated three structural patterns for managing iteration states within IBM Carbon v11:
I chose a responsive canvas grid using Carbon's 16-column layout system. It splits the editor into 2 or 3 synchronous panes. It uses more screen space, but it enables direct visual comparison and concurrent execution, which solves the core problem directly.
The proposed compare framework builds directly on existing Carbon v11 components. Composing natively within enterprise rules shows how to scale patterns across shared design systems.
Switches the single editor into 2 or 3 side-by-side panes, managed by Carbon's responsive layout grid.
A single Carbon Primary Button triggers every pane at once. Global parameters like temperature stay locked, ensuring fair comparisons.
Each pane's output, token count, and latency sit inside a Carbon Data Table directly below its prompt for instant evaluation.
Each pane contains its own prompt editor and filterable model Dropdown, allowing users to test different foundational models side-by-side.
Users select the winning variation to collapse back into single-editor mode with the victorious prompt and model parameters loaded.
Enterprise product work requires strict alignment with established component libraries rather than introducing custom UI patterns. Every component in this proposal is an unmodified Carbon instance.
Handles responsive 2-column or 3-column split layouts across desktop screen sizes without breaking editor proportions.
Uses Carbon Tabs to shift viewports between single-pane editing and multi-pane comparison modes.
The Dropdown component functions as the model selector inside each prompt pane, letting users swap models directly without leaving the workspace.
Formats inline output text, execution latency, and token consumption metrics cleanly.
Slides open to adjust shared temperature and top-p configurations across all active panes simultaneously.
Working inside an existing design system without breaking its rules requires technical discipline. It demonstrates the exact system-level thinking needed for multi-surface enterprise software.