Prompt Surfaces: Where Do Prompts Live?

Context

“Add a prompt box” is not a UX strategy. Where prompts live dictates adoption, latency, and error rates. Good placement respects task intent, risk tier, and available context—so people move faster with fewer reworks.

Four Prompt Surfaces (and When to Use Them)

1) Inline Prompts (in the flow)

Small, contextual inputs embedded where the work happens (compose fields, ticket forms, code editors).

Use when: You can pre-fill context from the page and the action is low/medium risk.
Wins: Lowest friction; highest adoption; easy to A/B copy.
Watchouts: Scope creep; users try “chatting.” Add guard text and char hints.

2) Side Panels (assist, don’t block)

A collapsible panel that shows suggestions, rationale, and evidence without covering the primary task.

Use when: You need to show why with citations, alternatives, or diffs.
Wins: Great for explainability and compare/contrast; supports confirm/override.
Watchouts: Can become a second app. Keep headers tight; show the top 2–3 actions.

3) Slash-Commands (power shortcuts)

Keyboard-first commands (e.g., /summarize, /draft) that prebind task context.

Use when: Pro users repeat the same micro-tasks and want speed + consistency.
Wins: Fast; learnable; telemetry-friendly.
Watchouts: Discoverability. Provide a cheat-sheet and fuzzy search.

4) Background Agents (zero surface—work shows up)

No prompt UI. Agents observe events and propose ready-to-apply changes (e.g., draft reply, PR notes).

Use when: Context is rich and stable; proposals must be reviewable not auto-applied.
Wins: Time-to-value; reduces initiation overhead.
Watchouts: Surprise factor. Always include source, rationale, and one-click revert.

Placement Framework

Task intent: Create, transform, decide, or explain? Match surface to the verb.
Risk tier: T0 (low) inline or slash; T1 (med) panel with confirm; T2 (high) background → review gate.
Context availability: If the system can pre-bind 80% of inputs, prefer background agent + review.
Latency budget: If SLA < 300ms, avoid blocking chat; stream in side panel.
Auditability: Surfaces that alter records must store rationale, evidence, and overrides.

UX Patterns That Travel

Prompt hints: 1–2 examples in gray text; disappear on focus.
“Show work” toggle: Inline → panel expands to reveal citations and version info.
Confirm/override pair: Primary: “Accept & apply.” Secondary: “Edit or explain.”
Fallbacks: When models refuse, offer rule-based templates or search.
Undo everywhere: Persistent “Revert” chip with diff preview.

Instrumentation (Micro-telemetry)

Surface used (inline/panel/slash/agent)
Time-to-first-action and time-to-decision
Edit distance and revert rate
Abandonment (opened panel, took no action)
Override reasons (dropdown w/ free-text)

Recommended Actions

Map top 5 tasks and assign a primary surface per task (one choice, not three).
Pre-bind context (record metadata, selection, role) before exposing any prompt field.
Ship a side-panel “explain” view that every surface can open with one click.
Standardize confirm/override controls and log reason codes for learning.
Review telemetry weekly to move tasks from inline → agent as confidence grows.

Common Pitfalls

Chat everywhere: Generic chat UIs become dumping grounds; task completion drops.
Surface sprawl: Multiple places to do the same thing → fragmentation.
No undo: Increases fear and slows adoption.
Missing evidence: Suggestions without citations reduce trust and reusability.

Quick Win Checklist

Add a universal explain panel with sources and version info.
Introduce two slash-commands for your most repeated transformations.
Make “Revert” universal and logged with reason.
Start a weekly dashboard: adoption, edit distance, abandonment by surface.

Closing

Prompts don’t belong in one place—they belong where they cut friction. Choose the surface that fits the task and risk today, instrument it, and graduate toward background agents as evidence builds.

Essay by OneMind Strata Team