Why you don't know what you use AI for
This pattern is drawn from Anthropic's Reflect feature announcement, published 9 July 2026, and the 4D AI Fluency Framework the dashboard uses to categorise usage patterns.
Ask a knowledge worker what they use AI for and they will give you a category — "writing help" or "research" or "summarising things." The answer is genuine but vague in the way that matters: it masks where most of the actual value is coming from, and where the gaps are. You cannot improve a habit you cannot see.
Claude's Reflect dashboard, launched in beta on 9 July 2026 for Free, Pro, and Max subscribers with Memory enabled, gives you the data to see it. It is not a productivity score. It is a usage mirror: topic breakdown by percentage, most-active day, peak working hour, and a natural-language summary of the task patterns Claude has observed over one, three, six, or twelve months.
The feature was designed with deliberate restraint. Ryn Linthicum, Anthropic's head of wellbeing policy, described the intent at launch: the team built it "with an eye toward how we can upskill people's usage of Claude, not in a way that encourages them to spend more time with it." Total time spent with Claude is absent from the current dashboard precisely because it is not the number worth optimising.
How to access Reflect
You need Claude Memory enabled first. If it is off, go to Settings → Privacy → Memory and switch it on. Memory works through a roughly 24-hour processing cycle: Claude reads your recent conversations and distils working patterns into a compressed profile stored on Anthropic's servers. Reflect draws on that profile.
Once Memory is active, open Settings → Reflect in the web or desktop app. Set the time window to three months. Shorter windows underweight your real patterns; longer windows dilute recent change. Three months is the calibration most useful for identifying a habit worth adjusting now.
What the dashboard shows
Reflect opens with a natural-language summary of your recent activity — the topics you engage with most, the types of tasks involved, and the overall shape of how you work with Claude. Below that are specific figures: your most-active day of the week, your peak working hour, and your total chat count for the period. The topic breakdown follows: a list of usage categories with percentage distributions.
This last section is the most useful input for improvement. Reflect categorises your interactions using Anthropic's 4D AI Fluency Framework:
Delegation — handing off complete tasks with clear scope and outputs defined.
Description — giving accurate, specific context and instructions before the task begins.
Discernment — knowing when to use AI versus doing the work yourself.
Diligence — checking AI output before it becomes real work, a message, or a decision.
The breakdown shows which dimensions appear most in your conversations — and, by implication, which are absent. A knowledge worker who delegates heavily but has almost no diligence signal in their pattern has identified a gap worth addressing before anything else.
Copy this prompt
Review the Reflect dashboard, note your top categories and the figures, then run this prompt immediately after. You can type the categories rather than copying them verbatim — the model only needs the substance.
I just reviewed my Claude Reflect dashboard for the past three months.
My top usage categories and approximate percentages are:
[Paste or type your top 3–5 categories and their percentages]
My most-active day is [day]. My peak hour is [hour].
Review this against the 4D AI Fluency Framework:
- Delegation: handing off complete tasks with clear scope
- Description: giving accurate, specific context and instructions
- Discernment: knowing when to use AI vs. doing the work yourself
- Diligence: checking AI output before it becomes real work
For each of the four dimensions:
1. Based on my category breakdown, am I exercising or neglecting this skill?
2. Is there a task type visible in my data that I should be handling differently?
Then give me:
- The one dimension where my usage pattern most suggests a gap
- One specific change I could make this week to strengthen it
- A concrete example of what that change looks like for a task in my top categoryWhy each part earns its place
Pasting the percentages. The framework analysis is only useful if it is anchored to your actual usage, not a description of what you think your usage is. The numbers from Reflect force precision; without them the output is generic advice about AI in general.
Including day and hour. Timing data is a proxy for task type. A knowledge worker whose peak is Monday morning is probably using Claude for planning and synthesis. Someone whose peak is Thursday afternoon is more likely doing document review or stakeholder preparation. The model uses this to contextualise the category breakdown and produce more grounded suggestions.
The four-question structure. Asking about each dimension separately prevents the model from collapsing the analysis into a single headline. Delegation and discernment are often inversely correlated: a high-delegation pattern with low discernment is the riskiest combination — a lot is being handed off but little is being questioned. The structure surfaces this before it shows up in a mistake.
One change, one concrete example. Reflect audits that produce a list of improvements change nothing. The last two instructions — one change, one example for your top category — force the output to land somewhere specific rather than hovering at the level of principle.
What this will not fix
Reflect requires Memory to be on, and Memory requires a processing lag of roughly 24 hours. If you enabled Memory today, your dashboard will be thin. Wait one week of normal working before running the audit.
The topic categories in the dashboard are Claude's interpretation of your conversations, not a verbatim log. They can mis-categorise tasks that cross multiple domains — a brief written inside a research session might be filed under research rather than writing. Read the breakdown as signal, not as a precise inventory.
The prompt will not tell you whether your current AI usage is producing good outcomes. Reflect shows what you are doing; it does not measure how well it is working. The audit finds gaps in the pattern. Verifying whether your output is actually better after making a change requires a separate review step — run one three months later and compare the two breakdowns directly.
Finally, the 4D framework is Anthropic's categorisation, not a universal standard. If your work sits in a domain — legal, clinical, financial — where discernment and diligence carry higher stakes than in general knowledge work, weight those dimensions accordingly when interpreting the output.
Intuition about AI usage is unreliable. Running the Reflect audit once gives you data. Running it quarterly gives you a change curve: evidence that the adjustment you made three months ago is now visible in the pattern, and a new gap has appeared at the edge of what you have learned to delegate.