FitWhen to use it — and when not
Use it when
- First run and empty conversations
- After an answer, to suggest the obvious next step
- Products with a narrow, specific job (support, analytics, legal)
Skip it when
- Expert tools where suggestions are noise after week one
- When you cannot make suggestions specific to the user's data
AnatomyThe parts of the pattern
- Starter chips3–4 concrete tasks phrased as the user would type them.
- ComposerChips fill the box rather than sending instantly, so the user can tweak.
- Follow-up chipsGenerated from the last answer: shorter, deeper, next action.
- RefreshA way to see different suggestions without scrolling.
GuidelinesDo & don’t
Do
- Write suggestions as real tasks with real nouns: "Summarize this contract's risks".
- Personalise from context: the open document, the user's role, recent work.
- Let a chip pre-fill the composer so people can edit before sending.
Don’t
- Generic filler like "Tell me a joke" in a B2B tool.
- More than 4 chips — it becomes a menu nobody reads.
- Keep showing starters to power users who never click them.
In the wildReal-world examples
ChatGPTClaudeGeminiPerplexity (related questions)
Products named for reference only — no affiliation, and the demo above is an original illustration, not a copy of their UI.
For engineersImplementation notes
- Generate follow-ups in the same call as the answer (a small JSON field) to avoid a second round-trip.
- Log chip impressions vs clicks per suggestion — it is the cheapest prompt-discovery analytics you will get.
- Cache starter sets per role/workspace; regenerate on a schedule, not on every page load.