Chat is not always the answer: when a button beats a prompt box

Picture a photo app that adds AI. The team ships a text box that says "Ask me anything about your photo." Most people stare at it, type "make it better", get something odd back, and close it. Meanwhile, the one thing they actually wanted was to remove the background, which could have been a single button.

Chat feels like the obvious way to add AI to a product, because that is how most of us first met it. But a prompt box is one kind of interface among many, and for a lot of everyday jobs it is the slowest and most confusing one. This article gives you a way to decide, feature by feature, whether to give people a box to type in or a control to press.

What a prompt box asks of people

A prompt box is a blank text field where the person describes what they want in their own words. It is powerful because it can accept almost anything. It is hard for the same reason: it gives no hint of what it can do.

Two old ideas from usability explain why.

Recognition is easier than recall. One of Jakob Nielsen's ten usability heuristics is "recognition rather than recall": minimise the user's memory load by making elements, actions and options visible. A button labelled "Remove background" lets people recognise what they want. A blank box makes them recall it, and then put it into words.

Writing a clear request is a skill. In June 2023 Nielsen described generative AI as a new interaction style he called intent-based outcome specification. In his words, "the user no longer tells the computer what to do. Rather, the user tells the computer what outcome they want." He also wrote that current chat tools "have deep-rooted usability problems", and argued that much of the population is not articulate enough in writing to get good results from prompt-only interfaces. He expected future AI products to mix this style with the clickable controls we already know.

The designer Amelia Wattenberger made a similar case in her essay Why Chatbots Are Not the Future of Interfaces. Her test is simple: "Good tools make it clear how they should be used." An empty chat box does not. She also points out that "everything you put in a prompt is a piece of context", which means people have to type, every time, information the product often already has.

When a button wins

A button (or a menu, toggle, slider or chip) beats a prompt box when most of these are true:

• The job is common and well defined. Remove background, translate to Spanish, shorten this paragraph, summarise this thread. If you can name it in two to four words, it can be a button.
• The product already knows the context. The photo is selected. The email is open. The person should not have to describe what is already on screen.
• The options are a small, known set. Tone: friendly, formal, short. Length: shorter, longer. A set of three to five choices is faster to pick from than to describe.
• The result needs adjusting, not rewriting. "A bit warmer" is a slider. Typing "make the colours 10% warmer but not too orange" is work.
• People will do it again and again. A repeated task should become one click. Retyping the same prompt every day is a sign a button is missing.
• Mistakes are costly. A clear, labelled action with a preview is safer than free text that might be read the wrong way.

When a prompt box wins

Chat earns its place in other situations:

• The request is open-ended or new each time. "Draft a reply that politely says no, but offers next month instead" has too many parts for a menu.
• People are exploring, not executing. Brainstorming, asking questions about a long document, or learning a topic.
• The space of possible requests is too large to list. Data questions like "which region grew fastest last quarter?" are a good fit.
• Expert users want a shortcut. Some people prefer typing to hunting through menus, much like keyboard shortcuts.

Even then, the box works better when it is not blank. Suggested prompts, examples under the field and a short line about what the assistant can and cannot do all help people get started.

A simple test: the four questions

Before you design an AI feature, answer these four questions. Score one point for each "yes".

3 or 4 points: build a button, menu or slider. 2 points: a button with an optional "Customise" field. 0 or 1 point: a prompt box, with examples.

Treat this as a rule of thumb rather than a research finding, and test it with your users.

A worked example: a writing assistant inside an email app

Here is one feature, designed both ways.

Version A: prompt only

A sparkle icon opens a box: "How can I help with this email?" The person has to know the AI can change tone, guess how to phrase it, wait for a result, and then type again to adjust.

Version B: controls first, prompt as a backup

Version B covers the jobs most people will want with zero typing, and it shows them what the assistant is capable of. The open field is still there for the unusual request. It has stopped being the only door in.

Here is the same idea as a before-and-after for a single task:

• Prompt: The person types "can you make this less formal and a bit shorter, keep the meeting time." The AI might drop the meeting time.
• Buttons: The person taps "Friendly", then "Shorten". Each step is small and visible, and the result can be undone one step at a time.

Hybrid patterns that work well

You rarely need to choose only one. These patterns mix controls and chat:

1. Suggestion chips above the box. Three or four tappable starters ("Summarise", "List action items", "Draft a reply") that either run straight away or fill the box so the person can edit.
2. Context-aware buttons. Show "Explain this formula" only when a spreadsheet cell with a formula is selected. The button appears where and when it is useful.
3. Buttons that open a pre-filled prompt. Tap "Rewrite", and the box opens with "Rewrite this to be…" so people only finish the sentence.
4. Follow-up controls on the result. After a reply appears, offer "Shorter", "Longer", "Try again" and "Undo" instead of asking people to type their correction.
5. Promote frequent prompts to buttons. If many people type nearly the same request, that is your next button.