PTN—01
Prompt starters
Curated example prompts on the empty state that teach range and get people to a first success fast.
An empty text box is a test most people fail. Faced with “Ask me anything,” they ask the one thing they already know a search engine can do, get a mediocre answer, and conclude the product is a toy. Prompt starters replace that blank test with three to six concrete, clickable examples that demonstrate what the system is actually good at — and, just as importantly, what kind of language works.
The best starters are specimens, not slogans. “Summarize this contract and flag unusual clauses” teaches more in one line than a paragraph of marketing copy, because it shows the verb, the object, and the level of specificity the system rewards. Each starter should map to a capability you are confident in; a starter that produces a weak answer is worse than no starter at all, because you chose the demo and still lost.
Starters should be one tap from a real result. Clicking one should either run the prompt immediately or drop it into the composer fully formed and ready to edit. Anything that merely opens documentation or a tour squanders the moment — the person signaled willingness to try, and the fastest path to retention is a first output worth reading within seconds of that signal.
Rotate and contextualize. A fixed set goes stale for returning users and generic sets ignore what you know about the person. Claude and ChatGPT both vary suggestions across sessions; products with account context can do better, seeding starters from the person’s actual documents, code, or data so the first success is about their work, not a canned demo.
Anatomy
Good morning.
Ask about customers, invoices, or schedules — or start from one of these.
An empty-state composer seeded with clickable example prompts that demonstrate capability range.
- 1Composer. The open input stays primary — starters supplement it, they never replace it.
- 2Starter chips. Three to six concrete prompts, each phrased as a specimen of good input, not a category label.
- 3Capability spread. Each chip demonstrates a different strength — summarize, generate, analyze — so the set teaches range.
- 4One-tap fill. Clicking a chip populates the composer as editable text; the person can send it or make it theirs.
When to use
- The primary interface is an open-ended composer with no obvious first move.
- New users churn at the empty state or ask only trivial, search-shaped questions.
- The system’s strongest capabilities are not guessable from the UI chrome.
- You can generate starters from real user context (files, history, role).
Design considerations
- 01Write starters as complete, sendable prompts — a verb, an object, and enough specificity to model good input.
- 02Test every starter against the live model weekly; kill any that produce weak or wrong answers.
- 03Fill the composer with the starter text as editable content rather than firing it silently, so people learn the shape of a good prompt.
- 04Show a different capability per starter — five variations on summarization teach less than one each of five verbs.
- 05Personalize from real context when you have it: reference the person’s actual files, recent activity, or role.
- 06Cap the set at six; a wall of twenty suggestions recreates the paralysis it was meant to cure.
- 07Retire starters from the UI once someone has sent several prompts of their own — they have graduated.
Pitfalls
- ✕Starters written by marketing (“Unleash your creativity!”) that are category labels, not runnable prompts.
- ✕A starter that demos a capability the model handles badly — you picked the exhibit and it still failed.
- ✕Static sets that never rotate, so returning users see the same four chips for months and stop reading them.
- ✕Starters that occupy the visual center and bury the composer, signaling that free input is the advanced path.
In the wild
- Claude
- The new-chat empty state offers rotating example prompts spanning writing, analysis, and coding; clicking one starts the conversation with that prompt.
- ChatGPT
- Shows a small grid of varied suggestion chips on the empty screen, refreshed across sessions to cover different capability categories.
- Perplexity
- Surfaces trending and example queries beneath the search box, doubling as a demonstration of the kinds of questions it answers well.
- Notion AI
- The AI menu lists concrete actions like “Summarize” and “Improve writing” against the current page, so the first prompt is grounded in the person’s own document.