Stream of Thought

Named by Shape of AI. Exposing intermediate reasoning turns dead wait time into perceived progress and gives users a preview of how the answer is being built — errors become visible mid-flight instead of after. DeepSeek’s R1 made it a mainstream expectation almost overnight.

Why it works. Latency with narrative feels shorter than silence. And reasoning text doubles as a trust artifact: users learn whether the model understood the question before reading its conclusion.

Steal: collapse by default after completion; keep the thinking skimmable. Skip: raw unedited reasoning on high-stakes tasks — visible confusion erodes more trust than it builds.

Flow captures for this entry are in progress — screenshots land with the teardown that covers it. Entries never ship images we didn't capture ourselves.