For years, my Monday mornings have started the same way: open Google Play Console, scroll through Android Vitals for each app, and read the small waves in ANR rate and crash rate. With a single app this is barely a chore. With a portfolio of wallpaper and relaxation apps running in parallel — which I have been growing for twelve years as an indie developer — it is the heaviest hour of the week. About three weeks ago I shifted that round to Claude in Chrome. Not full automation. The reading and the small lookups go to the model; the decisions stay with me. So far the change has been quieter and more modest than I expected, and I want to record that quietness honestly.
Both of my grandfathers were miyadaiku, traditional shrine carpenters in Japan. As a child I watched them return to the same buildings for decades to check the joints and the wood. Looking at Android Vitals is, for me, a smaller version of that inspection ritual. I am not chasing growth in those moments; I am checking whether the parts I worked on are aging gracefully and whether a change of season — a new Android version — has introduced any quiet warping. That is the spirit I wanted to keep when I introduced AI into the loop: AI to prepare the reading, me to do the judging.
Why I chose Claude in Chrome specifically
I did consider pulling the numbers via the Google Play Developer API and turning them into a digest somewhere else. I decided against it for one reason: the Vitals page in Play Console carries context that disappears when you extract the numbers in isolation. The axis of the chart, the colour of a warning, the threshold badge, the device-model breakdown — they all carry meaning that is hard to capture in a raw CSV.
Claude in Chrome sits next to the tab I was already opening anyway. It reads the page in my session, with my permissions, while I am looking at the same page. I do not have to screenshot anything or paste it into a different surface. I can simply say, "for this app, summarise where the ANR rate is coming from by device, and compare it to the last two weeks." The model reads the page in front of it and answers. The footprint is small, but the psychological footprint on a Monday morning is large.
How the weekly review is structured
The review runs on Monday morning in a one-hour slot. I open Claude in Chrome, hand it the list of apps in Markdown, and have it walk the same pages in the same order every week. I keep a short prompt template, which is roughly this shape:
You are my operations review assistant. For each app, walk Google Play Console
in the order below and return only "delta vs last week" and "candidate next moves",
as bullet points. Max three lines per app. No commentary, just the numeric changes
and anything that catches your eye.
Walk order:
1. Dashboard (installs, uninstalls, revenue)
2. Android Vitals -> core vitals
3. Android Vitals -> ANR & crash
4. Android Vitals -> startup time & frame rendering
5. Reviews -> 1-2 star
6. Policy -> warnings / action required
The key clause is "no commentary." When an AI moves into the supporting role, it is easy for it to puff up its output with interpretation, which then doubles the reading load on the operator. Stating "deltas and candidate next moves only" up front keeps the output quieter and easier to scan.
What has worked well in three weeks
The first week I was sceptical. By the third week the rhythm has settled. Three things stand out.
First, the time cost of the Monday review dropped noticeably. Before, I was spending five to seven minutes per app jumping between tabs. Now Claude in Chrome reads first, I scan the summarised deltas, and I only open the original page when something flags my attention. The model surfaces "what changed," I decide "what to do." That split has been the single biggest improvement.
Second, I am catching metrics that I used to skim past. My eye habitually jumped to ANR and crash, but Claude in Chrome flatly reports lines like "startup time degraded 8% week over week" or "crash rate concentrated on Android 15 only." Device and OS breakdowns surface as deltas, which corrects for the blind spots in my own attention. There is something useful about a reader that does not share my biases.
Third, and perhaps most importantly, I no longer take the first emotional hit from a one-star review. When low-rating reviews land, the wording can move you before you read it. Asking Claude in Chrome to "classify last week's low-rating reviews into feature requests, bug reports, misunderstandings, or other, and just report the counts" lets me approach the same information as a number first, then read the original wording afterwards with a steadier head. I still write the replies in my own voice. Outsourcing only the triage step turned out to be both psychologically and operationally sensible.
What I deliberately kept off the AI
Three weeks in, I am also clearer about what does not belong in this loop.
The first is interpretation of policy warnings. Play Store policy notices are sensitive to small wording differences and have outsized consequences if misread. I let Claude in Chrome tell me whether a warning is present, and I read the original text myself. The cost of a misreading is in a different class from the cost of misreading a Vitals trend.
The second is the release decision itself. When Vitals worsen, deciding whether to roll back, hold for the next hotfix, or throttle the rollout depends on the current user base, the revenue band, the load on neighbouring apps, and the rhythm of my own week. AI can carry me up to the edge of that decision, but the last step belongs to me. With a cumulative portfolio that crossed fifty million downloads, the responsibility of the final call is something I do not want to delegate, even to a capable model.
Three operational notes worth writing down
Three small habits made the difference between a fragile setup and a usable one.
Hand over the app list every week. Claude in Chrome holds context within a conversation, but explicitly passing the walk list each time keeps the output stable. I send it as a Markdown checklist and ask it to return checkmarks once each app is reviewed. Missed apps drop sharply.
Hand over two weeks of prior numbers, not one. If you only show this week's value, the summary drifts towards abstraction. Pasting last week and the week before as a short table keeps the model anchored to specific numbers.
Keep a one-file note after each review. I write notes/android-vitals/YYYY-MM-DD.md containing the model's summary and the decisions I actually took. The note is less for the AI and more for the version of me that needs to remember three months later why I chose what I chose.
What I plan to try next
Next week I will start folding the revenue tab and the monthly AdMob report into the same walk. Vitals can be quiet while revenue undergoes structural change — mediation allocations tilt, CPM drops in a specific region — and watching that on a separate tab is becoming inefficient. I want to ask Claude in Chrome to walk Vitals, then revenue, then ads, and give me three lines on the correlation between them.
The other piece I want to add is auto-routing the classified low-rating reviews into GitHub Issues. Right now they live in my private notes, which makes follow-up slip. Having Claude in Chrome open the issue with the original review text and the classification feels like a small but worthwhile next step.
The Monday Android Vitals round was, honestly, one of the quieter chores of twelve years of indie development. Being able to keep the inspection ritual but hand the reading to AI — and keep the judgement for myself — is the use of Claude in Chrome that has felt most right to me so far. If you are running several apps in parallel, perhaps a version of this rhythm will fit your week too. Thanks for reading.