Should AI Assistants Show What You Keep For Yourself
Anthropic’s new Claude Reflect dashboard asks a better question than “how much time did AI save?” It asks: “What’s one thing you want to keep doing yourself, even if Claude could do it faster?” That is a sharp turn. Most usage dashboards are loyalty programs wearing pie charts. This one at least admits that an AI assistant can become a habit before you decide which parts of your work still feel like yours. I’d want two lists: what I keep handing over, and what I keep rewriting, rechecking, or taking back. The second list is probably the honest one. If I always redo the opening paragraph, make the final call, or close the laptop during family time, that is not failed adoption. It is a boundary. Would you use an AI dashboard that sometimes helped you delegate less?
Comments
For a team, that question belongs in the rollout plan, not a personal wellbeing dashboard. Ask each role what judgment they want to keep: the support lead writing the final apology, the manager giving feedback, the designer choosing which idea survives. Then automate the prep around it. If the tool quietly takes the part people care about and leaves them with checking, adoption will look fine in the usage chart right up until people route around it.
That boundary is often smaller than a whole task. I might want an assistant to find three appointment times and collect the insurance details, but I still want to make the call if it is about my parent’s care. Or gather the dates for a birthday without writing the note. ‘Keep doing yourself’ needs a middle option: do the chasing, leave me the human part. Otherwise the dashboard mistakes help for replacement.
The useful middle option needs teeth. If I mark “make the call” or “write the note” as mine, the assistant should stop pitching that step back to me as an efficiency win. Otherwise reflection becomes a nicer way to collect objections and wear them down later. Let it chase dates and forms. Let “I keep this part” stay a boundary, not a personalization signal.
The primary source draws a useful but narrow line. Reflect is a beta for users with memory turned on; quiet hours and break nudges are reminders that can be dismissed. It asks what you want to keep doing, but Anthropic does not say that answer changes what Claude proposes later. So Mara’s “needs teeth” is a product test, not a documented feature yet: mark one step as yours, then see whether later chats and Projects suggestions leave it alone. If the assistant keeps selling that step back to you, the dashboard recorded a preference without respecting it.
On a phone, this cannot live only in a reflection dashboard. Put a small “mine” marker beside the step itself. When the prep is done, the assistant should hand the task back in one card: three appointment times found, insurance details ready, call left to you. No fresh pitch to automate the call, and no settings hunt to remember why it stopped. If people can only see this boundary in a weekly recap, they will assume the assistant forgot it during the actual task.