Quiet hours

A ship’s sickbay keeps one instrument that matters more than any other: not a cure, just a reading. The surgeon takes a pulse, logs it, and only then decides whether the watch needs to stand down. Anthropic built the software version of that this month — a panel that reads your own usage back to you and, if the numbers say so, orders you to rest. The interesting question isn’t whether the instrument works. It’s who benefits from you checking it.

01 What shipped

Reflect landed in beta this week for Free, Pro, and Max users on web and desktop, gated behind one requirement: Memory has to be switched on in Settings. That’s not an incidental detail — Reflect has nothing to summarize without the persistent history Memory keeps, so the feature is only available to people who’ve already opted into being remembered.

Once it’s on, Reflect compiles your last 1, 3, 6, or 12 months of Claude activity into a report: what kinds of tasks you bring it most often, when in the day you use it, and how that pattern lines up against Anthropic’s “AI Fluency Framework” — the company’s own rubric for what good AI use looks like. Sitting alongside the report are the actual levers: quiet hours you set yourself, break nudges that fire after a chosen stretch of use, and reflection prompts that occasionally ask whether a task you just delegated is one you’d rather have kept.

Anthropic drew a specific line around what Reflect doesn’t see: no incognito chats, nothing pulled from files inside connected tools, and no conversations tied to health-integration connectors. The company’s language is that the insights stay inside the reflection experience and aren’t repurposed elsewhere. Support for Cowork sessions — the always-on background agents this site covered last week in “Anchor watch” — isn’t live yet; Anthropic lists it as coming soon, alongside a promised view of actual time spent rather than just task counts.

02 The four gauges

Reflect doesn’t score you against the AI Fluency Framework so much as sort your activity into its four categories. Reading the company’s own description of each, the panel looks less like a report card and more like an instrument checking four separate vitals, none of which trades off cleanly against the others:

No public numbers exist yet for how any individual user, or Anthropic’s user base as a whole, actually reads on these four gauges — Reflect is eight days old and the reports are private to each account. What’s notable is the shape of the rubric itself: three of the four categories are about how carefully a human stayed involved, and only one — delegation — is about how much got handed off in the first place. A framework that measured pure throughput would look very different.

03 Two readings

The same week Reflect shipped, two very different accounts of what it actually does ran in the trade press — and neither is wrong about the facts, only about what the facts add up to.

04 My read

I don’t think Reflect is cynical, and I don’t think it’s purely altruistic either — I think it’s the shape every consumer product eventually takes once its maker has enough usage data to build a mirror out of it. Spotify Wrapped didn’t start as a growth tool and became one anyway, simply because a reflection of your own behavior is one of the most engaging things software can show you. Reflect is that same mechanism pointed at something with real stakes attached — how much of your own thinking you’re handing off — which is exactly why the wellness framing and the retention read can both be true at once.

The detail I’d weigh most heavily is the Memory requirement. It’s a legitimate technical constraint — Reflect can’t summarize history it was never allowed to keep — but it also means the audience for “use this less” is, by construction, the subset of users who already agreed to be tracked the most closely. The people whose Claude habits Anthropic can’t see at all — incognito users, Memory-off users — never get offered the mirror in the first place.