PTN—04
Suggested actions
System-proposed next steps offered mid-flow, turning each response into a fork the person can take with one tap.
Prompt starters solve the first move; suggested actions solve every move after it. Once a response lands, the person faces a smaller version of the blank-box problem: what can I do with this? Suggested actions answer by proposing concrete follow-ups — “make it shorter,” “turn this into a table,” “draft the reply” — derived from what just happened, not from a static menu.
The pattern works because it converts recall into recognition. Most people will never learn that they can ask for a different tone, a different format, or a deeper pass; a chip that offers exactly that teaches the capability at the moment it is useful, and each tap is also a prompting lesson. Products that do this well see conversations deepen — the second and third turns are where AI stops being a search box and starts being a collaborator.
Relevance is the whole game. Suggestions computed from the actual response content feel like a sharp colleague; generic suggestions bolted onto every message (“Tell me more!”) feel like an upsell and train people to ignore the entire region. Better three suggestions that are right than six that are plausible — and better none than three fillers, because the empty state at least does no damage to trust.
When to use
- Conversations stall after one exchange because people don’t know what to ask next.
- A response has natural next operations — reformat, expand, apply, send — that free text would express clumsily.
- You want to teach capabilities (tone control, export, tools) without a tutorial.
- Downstream actions in your product (create the ticket, schedule the event) can be offered directly from the answer.
Design considerations
- 01Derive suggestions from the actual response — its content type, length, and obvious next operations — not from a fixed list.
- 02Phrase each suggestion as the prompt it stands for, so tapping it is also a lesson in what to type next time.
- 03Cap at three; suggestion rows past that stop being read and start being wallpaper.
- 04Prefer actions that complete work (“Create these as Linear issues”) over ones that merely continue talking.
- 05Suppress suggestions entirely when nothing strong applies — a blank row outperforms filler.
- 06Keep them visually subordinate to the composer and dismissible without ceremony.
Pitfalls
- ✕Generic follow-ups (“Anything else?”) on every message, teaching users the region is decorative.
- ✕Suggestions that fire consequential actions — sending, publishing, deleting — from a single ambiguous tap.
- ✕Engagement-bait suggestions optimized to extend the session rather than finish the task.
- ✕Suggestion rows that push the composer below the fold, taxing exactly the users who know what they want.
In the wild
- Gmail
- Smart Reply offers three short, thread-derived responses beneath each email — the canonical proof that recognition beats composition for routine moves.
- Perplexity
- Appends related follow-up questions after each answer, generated from the answer’s content, which reliably pull one-shot queries into deeper sessions.
- Notion AI
- After generating text, offers next operations — continue writing, make longer, change tone — as a menu scoped to the draft it just produced.
- GitHub Copilot
- Copilot Chat surfaces slash commands and contextual follow-ups like /fix and /tests tied to the code under discussion.