Overview
How might we design progress steps so people can trust and act on AI output?
When to use
- Use when the system performs visible intermediate work (search, tools, multi-step reasoning) and you need both a calm default UI and optional transparency for verification, learning, or troubleshooting.
When to skip
- Single-step answers with no intermediate tool use or reasoning worth surfacing.
- Latency-critical UIs where rendering a step-by-step trace adds visual overhead to a near-instant response.
- Audiences who only care about the final answer and never want to see the underlying process.
Rules
Expanding the full trace by default, burying the actual answer under process detail.
Steps with no timestamps or duration, making it impossible to tell what took long.
Generic step labels like "Processing" that give no real insight into what happened.
Traces that disappear once the response finishes, preventing later review.
Evidence
| Product | Implementation |
|---|---|
| Perplexity | Collapsible "Steps" trace shows search queries and sources visited during research. |
| ChatGPT | Reasoning and tool-call summaries collapse under a "Thought for Ns" toggle. |
| Claude | Extended thinking renders as a collapsible block above the final response. |
| Cursor Agent | Tool calls and file reads appear as an expandable timeline during agent runs. |
