Lee Sun-woo

Let readers act before the summary is ready

In a local Pulse implementation from September 23, 2026, I let readers open a newly collected source while its summary was still being prepared. The article’s title, source and link were already available. Waiting for generation should not be a prerequisite for opening them.

This changed the question from how to decorate the delay to which parts of the reading task actually depended on AI.

Show what is ready without implying that everything is ready

I separated recent source links from the completed summary feed. A pending item could show the original title and a route to the article, with an explicit indication that processing was unfinished. It did not receive the same presentation as a completed Korean summary.

That distinction matters because the two items make different promises. One offers access to a publisher’s article. The other offers Pulse’s generated account of it. Reusing the completed card without a clear state difference would make unfinished processing harder to recognize.

Conceptual paths. Readers can open the source before its summary is ready.

The pending area also has a cost: it adds another place to look and may expose an untranslated title. For a reader who wants the Korean summary, waiting remains reasonable. The source link provides an additional route for someone whose goal can already be met.

Decide which failures should interrupt reading

A failed request for newly discovered items did not block the existing feed. The already readable material remained available. This follows the same dependency decision: completed stories do not depend on the next discovery request succeeding.

Anthropic’s Building effective agents discusses the latency and cost of adding model-driven steps. My design application is to inspect those dependencies before adding a waiting state. Opening an original article needs its URL; reading a generated summary needs generation to finish.

The transition into a completed summary still needs deliberate treatment. Could a story appear twice? Would inserting it move text someone is reading? These are review questions for this design, not outcomes established by the local implementation.

A useful waiting experience starts with an inventory of what is already available. Show that material honestly, preserve ongoing work, and reserve a completion promise for something the system can actually confirm.