interaction-speed

Interaction Speed

Perceived performance as a design property. The bar: snappy — the UI moves as fast as the user's thought, acknowledges every input instantly, and does its slow work where the user can't feel it.

Use This Skill When

• A flow feels sluggish, or users report clicking things twice
• Designing the responsiveness of a new interaction: what happens in the first 100ms, what's optimistic, what's preloaded
• Reviewing loading states, spinners, skeletons, and cache strategy
• Deciding whether an operation runs synchronously in the UI or in background

Example trigger phrasings: "make this feel faster", "why does this feel sluggish", "users double-click the save button", "should this be optimistic", "what should we preload".

Do Not Use This Skill For

• Diagnosing an actual performance regression to root cause — /th-engineering diagnosis (come back here for how the UI should behave while slow)
• Animation duration policy — visual-language

Reviewable Expectations

Latency budgets

Measure input → first visible response, not server time:

Budget	Requirement
< 100ms	Feels instant. Target for all local interactions: typing echo, toggles, menu open, palette filtering. No feedback needed beyond the state change itself.
100–300ms	Acknowledge immediately (pressed state, highlight), complete without progress UI.
300ms–1s	Show lightweight progress (inline skeleton/shimmer), keep the rest of the UI live.
> 1s	Run async: UI stays fully usable, progress is visible, completion is announced where the user is now — not where they were.

A hot-path interaction over budget is a defect regardless of what the backend costs; move the cost off the interaction.

Optimism and caching

• Optimistic rendering by default for mutations that almost always succeed: apply the change locally at once, reconcile in background, and on failure roll back loudly (restore state + explain), never silently. Never optimistic for irreversible or financial effects, actions with non-trivial failure rates, or server-authoritative results the user acts on next (generated IDs, balances, permission checks) — show honest pending state there; a rollback can't unfire a side effect.
• Preload what the user will probably want next: warm caches at idle, prefetch on hover/focus/route-likelihood, precompute the expensive view before it's opened.
• Stale-while-revalidate: render last-known data immediately, refresh in place. A blank screen that could have shown yesterday's data is a defect — except where stale data misleads or is unsafe to act on (balances, prices, permissions): there, block on fresh data or visibly mark the value stale.
• Cache invalidation states its rule (event, TTL, version) — "we'll refetch every time to be safe" is a latency defect wearing a correctness costume.

Input integrity

• Input is never blocked: typing, scrolling, and clicking stay live during background work. A frozen frame is worse than a slow result.
• Repeat-safe controls: the first activation disables, debounces, or dedupes; a double-click must never double-submit, double-charge, or double-navigate. A double-click on a single-action control is also a signal that feedback may have been late — investigate the feedback, don't assume.
• Fast users are honored, not throttled: type-ahead is buffered, an Enter pressed during a load is queued, keystrokes are never dropped while the UI catches up.
• Layout never shifts under a pending click: content that pops in must not move the button the user is aiming at.
• Context survives: navigation, refresh, and failure preserve typed input, selection, scroll position, and filter state. Destroying work-in-progress is a top-severity defect — worse than any latency finding.

Waiting and interruption

• Skeletons over spinners: show the shape of what's coming in place. Global spinners/overlays are a last resort.
• Progress is honest: no fake bars; if duration is unknown, say what's happening. Long-running CLI/TUI operations stream progress as they work; silent buffering until completion is a defect.
• Never make the user re-request: if it failed, offer retry with the input preserved.
• Interrupt (modal, focus-steal) only for decisions that must be made now; everything else is non-blocking and dismissible (toast, banner, badge), and never steals focus mid-task.
• Toasts carrying an action (undo, view) persist ≥5s and pause on hover; pure confirmations may auto-dismiss sooner. This skill owns interruption behavior; announcement semantics for assistive tech belong to accessibility.

Review Method

1. Walk the hot paths with the design-bar walkthrough's Pace and Repetition questions (design-bar); time input → visible response where measurable.
2. For each interaction over budget: classify the fix — acknowledge sooner, render optimistically, preload earlier, or move work async.
3. Grep mutation handlers for missing pending/disabled states and non-idempotent submits; cite file:line.

No built UI yet? Apply the same expectations to the spec's described interactions; evidence cites the spec section or screen region.