Browse all inspiration
The design inspiration
Inspiration
See how ChatGPT, Claude, Cursor, and others handle streaming, citations, approvals, and agent state — grounded in official docs and rebuilt as live demos.
Product studies
8 articlesThe streaming-first chat interface
OpenAI's ChatGPT defined the AI chat surface — streaming reveals, a composer that never leaves view, and citations that appear only once grounded.
Read breakdownReasoning traces, artifacts and a calmer visual language
Anthropic's Claude mainstreamed the thinking trace, artifacts that turn replies into canvases, and a warm low-contrast palette built for trust.
Read breakdownSearch and generation in one surface
Perplexity rebuilt search around an LLM — and its UI earns trust through visible tool calls, sentence-level citations and a persistent sources rail.
Read breakdownThe agentic editor that made agent state visible
Cursor turned coding into an agent task — and solved the trust problem with a persistent agent status, auditable tool calls and an explicit approval gate.
Read breakdownThe AI that stayed inside the document
Notion never built a chat destination. The model lives in the editor — space to draft, highlight to edit, accept or discard — and its answers cite the workspace pages they came from.
Read breakdownChat-first building with the code always in reach
Lovable builds apps from a conversation, but the interface keeps a way out of the chat: plan before build, visual edits without prompting, real code in a real editor, and a revert for everything.
Read breakdownThe assistant that keeps its drafts on the table
Gemini treats one answer as a starting point: alternate drafts to flip between, sources a click away, tools you can invoke by name, and a research plan you edit before it runs.
Read breakdownThe agent that keeps a save point under every step
Replit Agent builds whole apps from a chat, and the interface answers the fear that creates: checkpoints you can roll back to, a plan you approve before building, and a preview running beside the conversation.
Read breakdownDesign guides
4 articlesStreaming vs. waiting for the full reply
Most AI products stream by default — but streaming is a UX decision, not a protocol one. Here is when a streaming reveal earns its place, when a single rendered answer is better, and how to tell the difference.
Read guideWhen should an AI agent pause for approval?
The moment a tool call becomes consequential, an agent needs a human gate. Here is when to pause, what to show in the approval prompt, and how to keep approval from turning into a rubber stamp.
Read guideKeeping users oriented during long-running agent tasks
A task that runs for minutes needs more than a spinner: live progress, visible steps, and an escape hatch. Here is how the products that ship long-running agents keep users oriented — and in control.
Read guideInterrupting the agent: stop controls done right
Every long-running agent needs a way out — but a stop button that discards work is as bad as none. Here is how real products handle interruption: non-destructively, cleanly, and with the partial work kept.
Read guideMore coming soon — Notion AI, Gemini, Replit and Lovable case studies, plus more decision guides.