Draft mode lets users begin work using less details or less powerful processing power before committing to a full, resource-intensive run. This functionality supports an efficient, iterative process. Early generations go faster and users spend less compute or credits to get a preview, and users can easily switch to a higher-quality model once ready.
Versions of this pattern can be found across modalities:
While some products support an explicit draft mode, others offer a combination of parameters that creates a similar environment without being direct.
Visual mediums benefit from being specific, since a less powerful model will impact the overall quality of the generation. This guards user expectations and frames draft mode as a positive means of saving time and money in their creative process.
Implicit experiences across all content types allow users to decrease the step count of a generation, which reduces the fidelity of the AI's reasoning while drawing less compute power. This is a technical concept, though, so these options are often buried in settings for more savvy user to explore.
Decreasing the initial length or size of content also has the effect of creating within a drafting context. Audio and video generations, for example, often produce a short snippet of a high quality run for the user to review before expanding.
Similarly, text-based tools don't necessarily require an explicit mode since an outline is also a common writing tool and method of reducing the quality of the output before upscaling into a full document. Notably, text files use less compute power than other types of media, but making drafting mode explicit can still give users a sense of command over the amount of time spent. Jasper and other content editors mimic this choice by offering a parameter in the input box for users to choose to optimize for speed (draft) or quality (final).
As automatic model routing becomes more common, users increasingly rely on the system to select the best model for the task and iterative state. A form of drafting can be incorporated by defaulting to the less powerful alternative until instructed otherwise. This benefits the user by decreasing compute and credit consumption, but not all users will be wise to what is happening behind the scenes. However, less savvy users may be unaware that they can force the AI to use a more powerful model when needed.
Consider introducing this concept during onboarding in an interactive way, demonstrating that less powerful models can be the better tool earlier in the design process or while chatting about a topic before running a more powerful command like deep research.
