How should an AI product design the empty prompt box?

Illustrative situation: a founder ships an assistant that drafts replies to tenant messages for small landlords. Sign-ups are healthy. Then the logs show two groups. One closes the tab within a minute. The other types “help with lease” or “is this ok?”, with no file and no detail, and gets a polite request for more. The model is not failing. The box asks users to know what the product can do, how to phrase a request and what will come back, and gives them nothing to go on. The scene is invented, but the pattern is common.

Humbleteam, a digital product design agency, publishes this guide. Product behavior is described as of October 29, 2026. The guide treats the first input as four decisions: what the screen says about the product, what it offers to click, which parts of a request get structure instead of free text, and what the first reply does. A decision table and an invented worked example follow.

Why is a blank prompt box hard to use?

Jakob Nielsen calls the problem the articulation barrier: a prompt box demands clear written prose about a need the user may not have fully formed. His 2023 essay is an opinion piece. He bases his estimate of how many people struggle on reading-skill surveys and says he found no large-scale studies of writing ability. His NN/g article on intent-based interfaces also says chat interaction suffers from requiring users to write out their problems as prose.

A CHI 2024 paper, Bridging the Gulf of Envisioning, measures the distance between a person’s intention and a prompt that makes good use of the model. It splits that distance into a capability gap (setting goals the model can accomplish), an instruction gap (how to instruct it) and an intentionality gap (what to expect back).

A CHI 2023 study, Why Johnny Can’t Prompt, observed 10 people with little prompt-design experience. They explored opportunistically, not systematically. The authors gave them a baseline chatbot to improve, to avoid “Blank Page Syndrome,” where users facing a blank page do not know where to begin. Those people were designing chatbots, not using a product, so read the study as support, not proof.

For a product team, that means the box has to answer three questions before anyone types. Can this do my task? How do I phrase it? What will I get?

What should the screen say before the first prompt?

Microsoft’s HAX Toolkit opens with make clear what the system can do. It lists five patterns, among them an introductory blurb, demonstrating possible system inputs and showing a set of system outputs. The next guideline, make clear how well the system can do what it can do, says people often over- or under-estimate how many mistakes an AI system may make, and that unrealistic expectations can lead to disappointment and product abandonment.

Google’s People + AI Guidebook asks teams to be up-front about what the product can and cannot do the first time someone uses it, to explain the benefit rather than the technology, and to suggest a low-risk or reversible action to try right away.

For a prompt box, our suggestion is two plain lines outside the input: one naming the jobs, one naming a real limit. “Drafts replies to tenant messages and summarizes lease clauses. Does not give legal advice.” The wording is ours, not a Microsoft or Google format. What happens when a request falls outside that line is a separate problem, covered in when an AI product should ask, qualify or decline.

What do good starter prompts look like?

NN/g’s guide to use-case prompt suggestions, by Tim Neusesser and Kate Moran in June 2025, defines them as examples of good AI prompts, shown usually but not only to novices. Its advice, condensed:

• Match complexity to the tool. Short clickable pills suit broad tools and simple tasks. Cards suit specialized tools. Simple prompts can sit together, but complex ones are better shown one at a time.
• Prefer specific, situational prompts to generic ones.
• Curate by hand instead of auto-selecting from recent activity.
• Do not oversimplify. NN/g said Claude’s carousel examples arguably misrepresented the complexity of working with it.
• Allow one-click insertion.

The article gives no ideal count. Our rule of thumb is three or four, each a different job. Treat that number as a starting point to test.

In Microsoft Copilot, admins can pin up to four organizational prompts, which then appear as suggested prompts in Copilot Chat. Choosing a prompt in the prompt lab places it in the message box without submitting it, and matching prompts appear as typeahead while the user types. Google’s Gemini Enterprise uses prompt chips that act as categories: clicking one shows suggested prompts users can send immediately. In both, an administrator curates the set. A startup curates its own.

A good starter has a verb and a job, a slot for the user’s own content, realistic detail and a way to edit before sending. Notion’s help guide shows this in its sample prompts, which name the user’s own material: a launch-date decision, a PDF to turn into a one-page brief, a style-guide page referenced with @.

Should starter prompts be personalized or generic?

NN/g splits the answer by moment. Before login, suggestions work like marketing: a focused set of broadly relevant strengths, easy to dismiss. After login, they should be context-aware. It cites Tripadvisor’s assistant offering situational questions and Amazon’s Rufus adapting to the product on screen. It also says content should be tailored to the user’s level of experience and, without personal data, can draw on patterns from other users.

Our suggestion for an early product is three tiers: generic prompts by job on the first visit, adjusted by the role or goal the user stated after sign-up, then drawn from their own content after the first successful run. Personalizing on day one means guessing from nothing, and a wrong guess looks worse than a generic prompt.

Example inputs or example outputs?

HAX lists “demonstrate possible system inputs” and “show a set of system outputs” as separate patterns, and the difference matters. An input teaches phrasing. An output shows what the work looks like when it comes back. NN/g describes example libraries that pair prompts with results, common in image and video tools. In Midjourney’s library, selecting an image also shows the prompt behind it. Microsoft’s Prompt Gallery holds prompts, videos and articles, and each prompt carries tips for personalization and extension.

Our rule: show inputs when phrasing is the hard part, outputs when the doubt is whether the result is worth the effort, and both when producing the result is slow or costly. Generate sample outputs with the product itself, and regenerate them when the model changes. A tidy sample the product cannot repeat on a messy file is worse than none.

When does structure beat free text?

Nielsen’s best guess, in that 2023 essay, is that successful AI interfaces will be hybrids that mix prompts with graphical elements showing what can be done. His 2025 piece on prompt augmentation lists six patterns, including prompt builders, where dropdowns assemble a prompt, and parametrization, where sliders vary it. The CHI 2024 paper, after analyzing 12 systems, states a tenet: if tasks are well defined, people prefer dedicated interfaces over dynamic ones.

NN/g’s article on prompt controls, by Feifei Liu in August 2024, covers controls around the input. It saw Baidu’s Ernie show supported file types and size limits in a tooltip over the upload icon, and Perplexity’s Focus narrow the search domain. GitHub’s Copilot chat cheat sheet says slash commands help you avoid writing complex prompts for common scenarios. Users find them by typing a slash.

What should the placeholder say?

NN/g’s article on placeholders, by Katie Sherwin, concerns forms. It says placeholders vanish when typing starts, can be mistaken for default values and often have weak contrast, and that visible labels outside the field work better. A chat box is not quite a form, so what follows is our reading. Keep the capability line and the limits outside the box, and use the placeholder for one example at the right level of detail, such as “Paste a tenant message, or describe a clause question.” If nothing is lost when it disappears, it is safe.

How should the first response onboard the user?

The first reply is where users learn whether the product understood them. NN/g’s empty-state guidance, by Kate Kaplan in 2021, says empty states should communicate system status, provide learning cues and offer direct pathways for key tasks. It was written for application screens, but the same jobs fit the box and the reply.

Our suggestion for a first reply has four parts. Say what it used: “I read pages 1 to 12 and found the pet clause on page 4.” Deliver the result. Offer two next steps that are real follow-up jobs. If the request was thinner than the job needs, say which detail would improve it instead of refusing.

How do you measure whether the box works?

These definitions are our suggestion, not an industry standard. Set the window and the success rule per job.

• Box-to-first-prompt rate: new users who send a prompt in the first session, divided by new users who see the box. Its complement is abandonment before the first prompt.
• First-prompt success: first prompts where the user keeps, copies, exports or acts on the result, or where a reviewer scores the task as done.
• Reformulation rate: first prompts followed in the same session by a message that restates, narrows or corrects them. Exploratory tasks need some refining, so read it by job.
• Starter use: first prompts that began as a starter, split into sent as is and edited.
• Dead-end rate: first prompts that end in an out-of-scope or “need more detail” reply.

Clicks show curiosity, not success. NN/g notes that engagement is often driven by placement as well as content and suggests randomizing the order of starters. Microsoft’s admin analytics report active users and submissions per organizational prompt, which is a floor, not a success measure.

Illustrative worked example: a lease assistant

Every number and name here is invented. Wobblefern Rentals, a startup, sells an assistant that summarizes lease clauses and drafts replies to tenants. In one week of logs, 200 new users see the box. 58 leave without sending anything. Of the 142 first messages, 61 are under four words, and 37 refer to a lease that was never attached.

The team makes five changes:

1. Two lines above the box: “Summarizes lease clauses and drafts tenant replies. Does not give legal advice.”
2. Three starters, each a different job with a slot for the user’s content: summarize one clause in a lease, draft a reply to a tenant reporting a leak, compare notice periods in two leases. Each opens the upload picker or a paste field and can be edited before sending.
3. An attachment control that shows “PDF or Word, text-based, up to 20 MB” before upload.
4. A placeholder with one example of detail: “Paste a tenant message, or ask about a clause.”
5. A first reply that names the pages read, delivers the answer and offers two follow-ups.

For two weeks they log the five measures above and rotate the order of the starters. They claim no improvement, because they have not measured one.

Where do these rules stop working?

The sources mix HCI research, vendor documentation and practitioner guidance. NN/g’s observations of ChatGPT, Claude, Perplexity and Ernie come from its 2024 and 2025 reviews, and we did not re-check how those home screens look today. The Microsoft, Google, Notion and GitHub details come from documentation read on October 29, 2026, and interfaces change. We found no controlled study showing that starter prompts raise retention or first-prompt success, so test the advice here before you rely on it.

What can you produce now?

Five things you can produce before anyone redraws a screen:

1. A capability line and a limits line, approved by whoever knows the model’s real behavior.
2. Three or four starters, each checked against five questions: is it a job real users do, does it run correctly today, does it carry a slot for the user’s content, can it be edited before sending, and is the sample output current?
3. A row in the decision table for each job, naming the entry type.
4. A placeholder, a first-reply pattern and an empty-state variant for first visit, return visit and no data.
5. An event spec and a test plan for the five measures.

If you are choosing a team to design these states, our comparison of AI-native design agencies that show how they work on startup products shows which studios publish their process, so you can judge them before you talk.