llm-ui-patterns

Design UI for LLM-powered features — streaming response display, hallucination states, RAG source attribution, generative UI patterns, and AI trust calibration. Use when asked to "show streaming outpu

SKILL.md

GitHub
<!-- Adapted from phazurlabs/ux-ui-mastery (MIT) — LLM UI Patterns skill #17. Streaming response, hallucination states, RAG attribution, trust calibration. YAMTAM structure, Anti-Fake-Pass gate, and Vietnamese localization notes are original. -->

When to Use

  • Use when: building any UI that displays LLM-generated content
  • Use when: streaming responses feel jarring or incomplete-looking
  • Use when: users can't tell what AI generated vs what is factual
  • Use when: error states from LLM calls have no recovery UI
  • Do NOT use for: prompt design — use prompt-engineering
  • Do NOT use for: RAG retrieval architecture — use rag-architect

Streaming Response UI

Core principle

Show tokens as they arrive — never wait for complete response. Users perceive streaming as faster even when total time is equal.

Implementation pattern

States during streaming:
1. Thinking     → subtle pulsing indicator (not a spinner — implies processing)
2. Streaming    → text renders token by token; cursor blink at end
3. Complete     → cursor disappears; copy/action buttons appear
4. Error        → replace content with error state (preserve partial if useful)

Cursor pattern

/* Blinking cursor at end of streaming text */
.streaming-cursor::after {
  content: '▋';
  animation: blink 0.7s step-end infinite;
}
@keyframes blink {
  0%, 100% { opacity: 1; }
  50%       { opacity: 0; }
}
/* Remove when streaming completes */
.complete .streaming-cursor::after { display: none; }

Interrupted stream handling

  • Network drop mid-stream: show partial content + "Response interrupted" banner + retry
  • Never silently show partial content as complete
  • Preserve partial content for retry — don't discard

Hallucination and Uncertainty States

AI content needs uncertainty signals — users cannot judge correctness without them.

Confidence levels (display)

Confidence Visual treatment
High (cited, grounded) Normal text
Medium (inferred) Subtle italic or muted color
Low (speculative) Inline warning badge: "Not verified"
Unknown (uncited) Always label — never display as fact

Attribution pattern (RAG)

[AI response text here with a cited claim¹]

───
¹ Source: Employee Handbook, Section 4.2 (updated 2025-01-10)
  "Vacation accrues at 1.5 days per month after 90 days."

Rules:

  • Sources must link to the actual document/section, not just "database"
  • Show retrieval timestamp when data freshness matters
  • "Based on X documents" is not attribution — show which documents

Hallucination error state

When model explicitly signals low confidence or when retrieval returns nothing:

╔════════════════════════════════════════════╗
║  ⚠  I don't have reliable information      ║
║     about this.                            ║
║                                            ║
║  I searched 3 knowledge sources and found  ║
║  no matching content.                      ║
║                                            ║
║  [Search the web]  [Ask a human]           ║
╚════════════════════════════════════════════╝

Never display a confident-sounding hallucination. Surface uncertainty.


Generative UI Patterns

Progressive disclosure for long responses

Short response (< 200 words): show in full
Long response (> 200 words):  show first 100 words + "Show more"
Very long (> 1000 words):     summarize + collapsible sections

Regenerate and edit patterns

Every AI response needs:

  • Regenerate: "Try again" — same prompt, new generation
  • Edit prompt: let user refine query without full re-type
  • Copy: one-click copy of full response
  • Thumbs up/down: inline feedback for RLHF signal (if applicable)

Structured output rendering

When LLM returns structured data, render it — don't dump raw text:

LLM returns: {"items": [{"name": "Setup", "status": "done"}, ...]}
Render as:   ✓ Setup    ✗ Deploy    ○ Test
Not as:      {"items": [{"name": "Setup" ...

AI Trust Calibration

Users over-trust or under-trust AI outputs — design for the right level.

Over-trust signals (add friction)

  • Users submitting AI-generated legal/medical content without review
  • No human review step before high-stakes actions

Fix: Add explicit confirmation: "Review this AI-generated content before sending." Add edit affordance so users feel ownership, not passthrough.

Under-trust signals (reduce friction)

  • Users re-asking the same question 5 times with slight wording changes
  • Users copying AI output to verify in Google separately

Fix: Show sources inline. Show confidence. Make AI reasoning visible ("Here's why I think this...").

Trust vocabulary (copy guidelines)

Instead of Use
"The answer is..." "Based on [source], ..."
"Always..." "In most cases, ..."
"I know that..." "According to the docs, ..."
Error code dump Plain language + action

AI UI States (7-state mapping)

Map ui-states 7-state model onto AI-specific scenarios:

State AI-specific behavior
Empty "Ask me anything about [domain]" + suggested prompts
Loading/Thinking Pulsing indicator — not spinner
Streaming Token-by-token render + blinking cursor
Partial Show retrieved sources while generating
Complete Full response + copy/regenerate/feedback actions
Error Plain-language AI error + retry + escalation path
Offline Queue prompt, show "Will respond when reconnected"

Reference ui-states skill for implementation details per state.


Anti-Fake-Pass Rules

Before claiming an AI UI is done, you MUST show:

  • Streaming: thinking → streaming → complete → error states all implemented
  • Interrupted stream handled (not silently shown as complete)
  • Hallucination/uncertainty: at least "low confidence" visual treatment defined
  • RAG attribution: sources shown per claim, not just "AI-powered"
  • Regenerate + copy actions on every AI response
  • Trust-appropriate copy — no confident absolutes on uncertain outputs

Reference: gates/anti-fake-pass-gate.md | gates/ui-quality-gate.md