Capability hints must move with the dialogue; a one-shot start-screen hint does not cover later turns

What it is

“Try drafting an email” on the start screen helps before the first sentence. Ten turns later the user is revising a table or chasing a bug in code; that hint is still sitting there, or it vanished with the empty state. Either way it no longer covers what can be said now. Mid-dialogue capability hints have to update with topic, object, and the artefact already on screen. A one-shot first-screen hint is an opening line, not discoverability for the whole session.

The failure is not that the hint arrived late. It is that the hint is frozen in the wrong phase of the conversation. Once the phase moves, the old hint is noise or a blank.

Why it happens

Dialogue hides task state inside turns. When a graphical interface changes page, the toolbar changes with it. A conversation has no page; unless hints are actively swapped, the action set in the user’s head stays at the opening. People keep applying opening verbs to an object that has become a table, code, or a long draft, then blame their wording.

The context window already contains enough signal to know the current object (a draft, an image, a column of numbers). If hint generation runs only on session_start, that signal never enters the hint logic. The cheapest engineering choice is to cache the first-screen copy, which quietly turns discoverability off after turn one.

Studying it

Slice sessions into phases: opening, editable artefact present, after a topic shift, recovering from failure. At each phase, record the hint set on screen and code whether it names a legal next step on the current object. Independent variables: hint policy (freeze first screen / refresh every N turns / refresh after a detected topic shift). Dependent variables: phase-match rate, whether the first request after a shift still uses an opening verb, time to discover a new verb after the shift.

Do not operationalise topic shift with a keyword list — it misses cases. Use interface events: a new file was attached, output form changed from prose to table. In a retrospective interview, ask “what did you think you could still do in that turn” and compare with the hints actually shown; that catches “the hint was there, I treated it as expired.”

Where it stops holding

A one-and-done tool (translate a sentence and copy) has no later phase; the first-screen hint is the whole journey. If phases chop too finely, swapping chips every turn flickers and people stop reading; updates should bind to a change of object or task type, not to every sentence. In privacy-sensitive settings, generating hints from content can write a task the user did not mean to display onto the screen; fall back to generic verbs.

Applying it

• Give hints triggers: artefact type changed, a new attachment was pasted, two consecutive turns edited the same object. On trigger, replace chips rather than append to a lengthening list.
• Make chips speak to the current object (“add a totals column to this table”), not reuse the opening “write an email.”
• Remove stale hints from the input region; do not leave them greyed. An expired chip is worse than none, because it keeps acting as the wrong ceiling.
• Check: sample 20 sessions longer than six turns that changed object mid-way. Screenshot the hints by the field at each turn and mark whether they match the object of that turn. A mismatch lasting more than one turn is a defect. Then look at the first sentence after the object change: if it is still an opening verb forced onto the new object, the hints did not hand over the phase.

Related

• Same group: L2.09.1 Natural-language interfaces have nothing to scan, so discoverability is worse than in graphical ones · L2.09.2 Examples draw the range of what is possible, so diversity matters more than count · L2.09.3 People treat seen examples as a ceiling; a narrow set shrinks real use · L2.09.4 Suggestions that appear only after failure arrive past the point of giving up
• Nearby: L2.14 Carrying and Clearing Context · L2.02 Discoverability of What Can Be Said · L2.05 Iterative Revision
• Search terms: mid-dialogue capability hints · context-sensitive help · stale empty-state cues