Skip to content

Case studies ​

Each case links to public code at a fixed commit, so you can check how the package is used. Cases written by the package's author say so.

Amazing Claude Code ​

Written by the author of smooth-stream-text, who also builds this app.

Amazing Claude Code is an open-source chat panel for Claude Code inside JetBrains IDEs. The panel is a React 19 app in the IDE's embedded browser, and it renders answers with its own Markdown renderer.

Problem. The answer reaches the panel as the Claude Code CLI prints it: sometimes a few characters, sometimes dozens of words in one event, with pauses while the model works. Rendering each event as it arrived made the text jump forward in batches.

Integration. smooth-stream-text 0.2.0, two layers:

  • Pacing: one useSmoothStream call paces the answer that is currently streaming, with done following the CLI's stream (Feed.tsx, line 285). Finished answers are not paced at all.
  • Reveal: the panel's own Markdown renderer wraps each text run it produces in Reveal (Markdown.tsx, line 175), and one RevealProvider per message keeps the wave continuous across paragraphs (TextCard.tsx, line 121).

Pacing and reveal read the same unit setting ('char'), so the pace and the animation stop at the same places (TextCard.tsx, line 56). On touch screens the blur is turned off and only opacity animates.

Result and limits. The answer now flows at an even rate instead of arriving in batches. This was judged by eye in daily use; there are no before-and-after measurements for this app. For numbers on recorded streams, see the benchmarks.

Add yours ​

If your project uses smooth-stream-text and the code is public, open an issue or a pull request on GitHub with:

  1. The project, a link to its code, and who you are in relation to it.
  2. The problem: what streamed text looked like before.
  3. The stack: framework, renderer, transport, and the smooth-stream-text version.
  4. How it is wired, with links to the lines at a fixed commit.
  5. What changed, and how you know: a measurement, a recording, or "judged by eye".
  6. Anything that did not work well.

Only what you are happy to publish. Your own words about the result, good or bad, are what makes a case useful.

MIT licensed. Plain-text docs for tools and assistants: llms.txt, llms-full.txt, or append .md to any page URL.