Open input

Open input has become the cornerstone of interactive AI design, fostering a dialogue between users and AI systems. This pattern is characterized by its simple interface that feels familiar, inviting the user to converse with the model underneath.

By using natural language, it doesn't take long for someone to get comfortable with the general interactivity. Where the pattern's limitations show is after the first few interactions, when someone doesn't know what to say next.

Variations and forms

This action is versatile and appears in multiple contexts across AI experiences. Some parameters and other selectors may not be appropriate in all contexts so ensure you design each case thoughtfully.

  • Chat box. A persistent input at the bottom of a conversation. Best for back-and-forth tasks, discovery, and follow-ups.
  • Inline composer. A prompt that operates on a selection or cursor position inside an editor, such as when text is highlighted. Good for precise edits like Generative Fill and inpainting.
  • Command-style prompt with parameters. Single line prompt plus structured flags or controls, common in image models like Midjourney. Good for experts who want precision.
  • Side panel composer. A prompt in a panel that can pull in files, tools, and settings, used for longer tasks and multi-source grounding. Common in Workspace side panels and IDEs.

Supporting features

There's a false perception that simple means easy. When someone knows what they are looking for then this way of interacting with the model makes sense.  This could apply to use cases like a search portal, or customer support.

However, when someone reaches an open chat bar and doesn't know what they are looking for (content generation sites, ChatGPT, etc), it can lead them to feel crippled by the choices (the blank canvas effect).

On top of that, prompting skills are not widespread. Most users will not understand how to craft a prompt to get the result they have in their head.

Help users get the most out of open inputs to the AI through wayfinders that get them started and tuners that help them easily build their prompt.

  • Templates can help users craft better prompts without having the full skill set
  • Nudges to improve your prompt can show users what “better” looks like
  • Putting filters and parameters at the users’ fingertips can make this more complicated feature feel accessible

Design considerations

  • Set a clear default scope, and make scope switching a single-step action. For open inputs after the initial prompt, make the target of the action clear. This prevents accidental whole-document edits and allows users to precisely communicate their intent.
  • Handle limits and gaps with constructive guidance. Like inline errors in more analog experiences, avoid generic errors or silent fail states. For example, if the input is too long or underspecified, say what is missing and offer a fix.
  • Support novice prompters. Do not rely on users to be expert prompt or context engineers. Provide wayfinding tools to help users get started, like example galleries, and continue to support the flow of work with automated suggestions, follow ups to maintain an understanding of the user's intent, and templates for more complicated inputs.
  • Retain user control at all stages. Continue to offer parameters, model selection, modes, and other precision controls when open inputs continue in the flow of work or conversation. Do not limit these powerful features to the initial prompt alone.

Examples

Figma Make uses open text inputs combined with the “pointer” pattern to direct the conversation to a specific place on the canvas.
Open text is the most common input type, made popular initially by ChatGPT. Almost immediately, companies started adding complications to the open text approach, adding parameters, modes, and contextual attachments, shown here in Julius.ai
Lovable explicitly allows users to flip between build and chat mode, so a single open text input serves to direct the AI and communicate with it.
Open text inputs are used outside of conversations, such as for generating prompts to train the AI itself, seen here in Personal AI