Problem
A capable assistant is worth nothing if nobody opens it. It has to be discoverable inside an interface people already navigate by habit, without displacing a control they use daily and without a tour.
Solution
Put the affordance next to the object it acts on. An assist that rewrites a paragraph belongs at the text selection. One that summarises a document belongs in the document’s header. One that answers questions about a dataset belongs beside the dataset. Proximity does the explaining that an icon can’t, and it’s the difference between a control people understand on sight and one they have to be told about.
HAX guideline 7 calls this efficient invocation. The useful reading is that invocation cost includes the cost of finding the thing in the first place. A global button in a corner is cheap to build and expensive to use, because the viewer has to carry the object to the assistant instead of the assistant arriving at the object.
The sparkle deserves a paragraph of its own. It became the field’s universal marker in about two years, and its meaning has correspondingly collapsed: it now signals that something is AI rather than what that something does. A sparkle on its own is close to unlabelled. Pairing the icon with a verb costs a few pixels and converts a genre marker into a control. Where space forbids a label, the tooltip and the accessible name have to carry the whole burden, which makes them worth writing carefully rather than filling with the product’s name.
Ambient versus invoked is the other axis. An entry point that waits to be clicked respects the viewer’s attention and gets used less. One that offers itself proactively gets used more and interrupts. Guideline 3 is about exactly this timing, and the workable compromise is usually to stay quiet by default and appear on a signal the viewer generated, like a selection, a pause, or an empty state.
Persistence has a cost that’s easy to underweight. A permanent button occupies a slot in a toolbar forever, and toolbars are the most contested space in any mature product. Attaching the entry point to selection or context rather than to chrome avoids spending that slot, and it keeps the interface from acquiring a control that most sessions never touch.
Use when
The assistant is a feature inside a product people already use for something else.
Don’t use when
The assistant is the product. There the entry point is the whole screen, and a sparkle is decoration on a surface that’s nothing to compete with.
Trade-offs
Contextual entry points are discoverable at the moment of need and invisible the rest of the time, so nobody learns the feature exists until they happen to be in the right place. A persistent global button is always findable and always slightly in the way, and it teaches the wrong model by implying the assistant works the same everywhere. Proactive offers lift adoption and generate the resentment that ambient AI features attract. An icon everyone has adopted carries recognition and no meaning—a poor trade for a control that does something specific.
Checklist
- Is the entry point next to the object it acts on?
- Does it carry a verb, or only an icon?
- What’s its accessible name, and does it describe the action?
- Does it appear unprompted, and on what signal?
- Can it be dismissed, and does the dismissal stick?
- What permanent space does it occupy, and what did that space cost?
- Is there a keyboard path for people who never use the button?
- Does it appear in contexts where the assistant cannot help?
- Does it look the same across the product when it does different things?
- Can someone who does not want it turn it off?
Compare
Notion puts the entry at the cursor and on selection rather than in the chrome, so invocation happens where the writing is and no toolbar slot is spent. Figma attaches its equivalent to the canvas selection, where the object is a frame and the available actions follow from what is selected. Microsoft Copilot takes the opposite approach with a persistent ribbon presence across applications, buying consistency and constant visibility at the price of a permanent slot in the most contested toolbar in software. Raycast removes the surface entirely and makes the entry point a global keystroke. For anyone who learns it that’s the cheapest invocation there is. For everyone else it may as well not exist.
Related
Inline assist is the most common thing an entry point opens. Assistant sidebar is the heavier destination. Command menu is the keyboard-first alternative to a visible affordance. Prompt starters handle the moment after the entry point is used. Ghost text is the version that needs no entry point at all and carries the costs that come with that.