PTN—02
Contextual entry points
AI affordances placed at the moment of need inside existing workflows, instead of a separate destination.
A standalone chat tab asks people to leave their work, reconstruct its context in prose, and carry the answer back by hand. Contextual entry points invert this: the AI appears where the work already is — a button on a selection, an action in a command menu, an inline affordance in an editor — carrying the surrounding context with it. The person never explains what document they mean, because the entry point already knows.
This is the difference between AI as a destination and AI as a material. Destination AI gets used when someone remembers it exists; contextual AI gets used because it sits on the path they were already walking. The gap in adoption is not subtle. Features reachable from selection menus and command palettes see routine use; the same capability parked in a sidebar tab becomes a novelty visited twice.
Placement is an editorial act, and restraint is the discipline. Every entry point you add competes for attention with the work itself, and a sparkle icon on every surface reads as desperation, not capability. Put entry points where intent is legible — a selection, an error state, a blank field with a known purpose — and resist decorating every hover state with an AI affordance that fires a generic chat.
When to use
- The AI operates on objects that live elsewhere in your product — documents, code, designs, records.
- Users would otherwise have to copy context into a chat and paste results back out.
- Intent is inferable from location: a selection, a failing test, an empty description field.
- A capability is going unused because it lives in a separate tab nobody visits.
Design considerations
- 01Attach entry points to objects and selections, not to global chrome — the selection is the context.
- 02Pass everything the location knows (file, selection, cursor, error) into the request automatically; never make the person restate it.
- 03Return results in place — edit the selection, fill the field — rather than opening a chat that quotes the answer back.
- 04Register every AI action in the command palette with plain-verb names people would actually search.
- 05Keep the entry point one gesture from dismissal; an affordance that is hard to wave off becomes an irritant.
- 06Show which context was captured before running, so people can trust — and correct — what the AI sees.
Pitfalls
- ✕Sparkle-icon sprawl: an AI affordance on every surface, most of which open the same generic chat.
- ✕Entry points that discard their context and make the person re-explain what they had selected.
- ✕Popovers that cover the very text the person is trying to work with.
- ✕Contextual actions that route to a separate destination anyway, adding a step instead of removing one.
In the wild
- GitHub Copilot
- Inline completions appear at the cursor inside the editor — the entry point is the act of typing, with the surrounding code as automatic context.
- Notion AI
- AI actions are reachable from the block menu and text selection, operating on the selected content and writing results back into the page.
- Cursor
- Cmd+K opens inline edit scoped to the current selection or file, so the request carries code context without any pasting.
- Gmail
- Smart Reply and Help me write appear inside the compose flow, drawing on the thread being answered rather than requiring a separate assistant.
- Figma
- AI features surface through the actions menu and on selected layers, so generation and renaming operate on what the designer already has in hand.