共有:

Anthropic Claude — Reflect

Claude now shows you
how you use it.

Following the year-end wrap-up ritual of Claude Wrapped, Claude just added Reflect to the app itself. It isn't a one-off event — it's an always-open mirror of your usage. Prompt habits, dependence on the model, time-of-day patterns, all now visible. Baking "self-review" into the UI is a quiet design choice that changes how Claude gets operated.

AI Navigate Editorial2026.07.116 min read

REFLECT usage mirror prompt habits time-of-day mix dependence score use-case breakdown an always-on dashboard, not a one-off recap
FIG. Reflect surfaces four axes — prompt habits, time-of-day, dependence, use case — as an in-app dashboard.
01

What Changed

Wrapped happens once a year. Reflect is "whenever."

Same recap, different cadence — and that changes the meaning.

Over the past few years, AI vendors' "recap content" has followed the Spotify Wrapped playbook — a year-end marketing event. Anthropic's Claude Wrapped 2025 was in that mold. Reflect is deliberately not an extension of that. It's a different product.

Anthropic's scope for Reflect covers four axes: (1) breakdown of the categories you prompt in, (2) time-of-day usage distribution, (3) a "dependence" indicator — how often you offloaded questions you could have solved unaided, (4) trend against the previous month. It's an always-open dashboard, not a shareable card. It's positioned as an instrument panel for self-review and operational tuning, not a social-media artifact.

02

The Numbers

The four indicators Reflect shows

4 axes
habits, time, dependence, use case
Always-on
not once a year
In-app
no extra tooling to install
Claude Wrapped (previous)Reflect (new)
Annual eventAlways-open dashboard
Made to be sharedInstrument panel for self-review
Consumer-firstDoubles as a team-operations signal
Ends at "that's fun"Prompts "so, what next?"

03

Why It Matters

"How you used AI" is now part of the product

LLM products used to be one layer — input goes in, output comes out. Reflect inserts a second layer: usage visibility. ChatGPT, Gemini, and Copilot all have analytics tabs, but almost all of them are admin-facing usage metrics. "Let the user themselves review their own usage" is where Claude has stepped forward first.

The difference matters. Letting the person doing the operating see their own metrics is the shortest path to changing behavior. Since Claude Code rolled out, Anthropic has been steering toward "measure the human-AI collaboration from the human side too." Reflect reads as the next chapter of that thread.

04

Who It Hits

Impact by reader type

Same feature, opposite meaning depending on your seat.

Team operators

You can see per-person how Claude gets used — which tasks, when, at what rate. That surfaces unevenness inside the team and areas where training is missing. It also feeds directly into license optimization.

Individual users

A visible "AI dependence" score is a double-edged sword. It can trigger healthy self-review — or people just glance at it once. Light users are likely to open it once and never again.

Compliance & audit

Users seeing their own usage pattern is useful material for DLP training and acceptable-use policy work. Audit posture (log retention, cross-border transfer) needs to be checked before you rely on it.


Reflect makes the AI
show you how you used it.


05

What's Next

What comes next / what to do

01

Watch your Reflect for a week

For individuals and teams alike, a week is the right observation window. If your prompt-habit distribution is lopsided, that's a chance to templatize and share what works across the team.

02

Predict which vendor follows

OpenAI and Google are likely to add similar dashboards. But the definition of "dependence" will differ across vendors — read the metric definitions before comparing scores.

03

Get individual consent before team rollout

Reflect is user-facing today, but an admin view is a matter of time. Agree on how personal usage data will be handled before you roll this out at team scale.

06

The Counterpoint

The other side: measurement warps behavior

Visibility has a side effect. Showing someone a "dependence score" can create the very measurement artifact where a person deliberately avoids AI in cases where they should be using it — just to keep their number down. The Hawthorne effect from productivity measurement literature applies here directly. Taking Reflect's numbers too literally can lead to inverted behavior.

Second, the metric definitions live entirely inside Anthropic's discretion. What counts as "a question you could have solved unaided" is decided by Claude, and the criteria are opaque. Treat Reflect's numbers as directional — a tool for spotting trends — rather than as absolute truth. That's the healthy read.