Support at 100 users is a habit. Support at 40,000 users is a department. Pocket Clear does not have a department — it has two people, a Hermes agent that reads every email and review into a tally, and Claude Code that turns the tally into features and reply drafts. This is the exact pipeline, including the parts that go wrong.
I run Pocket Clear, an expense tracker for people who want to notice their money. We are past 40,000 users. Half of them are outside the US — reviews arrive in Spanish, Portuguese and Bulgarian. The team is two people.
We started this the way most small teams do. At 980 users I had an always-on assistant on a Mac mini in a closet triaging the inbox and drafting App Store review replies. It gave me two hours a day back. But by 40,000 users the problem had changed shape: it was not that replying took too long. It was that I was reading every message one at a time and losing the pattern.
I would fix the bug in front of me, reply, and move on. Six weeks later a third person would ask for the same thing and I would have no idea it was the third time.
Then there was the bigger miss. For a year I pitched Pocket Clear as “privacy-first.” When we finally counted what our 65 store reviews actually talked about, privacy was 3%. Simplicity was 34%. Building a money habit was 25%. “No ads” was 23%. Two reviews out of sixty-five mentioned the thing I led every pitch with. The data had been sitting in my inbox the whole time. I was just not reading it as data.
The first agent does not talk to anyone. It reads. I run Hermes Agent with the email integration pointed at the support inbox, plus the App Store and Google Play review feeds. Every morning it goes through everything new and, for each message:
Once a week it writes a research brief on top of the tally: what is trending, which requests just crossed three votes, which market the requests are coming from. That last part mattered more than I expected. One request for three-decimal currency support means nothing. Three of them, all from Gulf countries, in one month, is a market telling you something.
The tally is a markdown file in our repo. 114 rows so far. 40 of them shipped. It replaced our roadmap — not “informed” it, replaced it. When a request crosses three votes it gets built. When it ships, every person who asked gets an email that says “you asked, it’s live.”
A few real rows:
Each of them got the “it’s live” email. And this is the part I did not plan for: 11% of our reviews now mention that they emailed us and the thing got built. The research loop became a retention loop.
Want this pipeline on your inbox?
We set it up as a fixed-price package — Starter from $1,000 for one inbox and one flow, Growth from $10,000 for the full inbox + reviews → tally → Claude Code loop.
The brief and the tally land in the repo. Claude Code takes it from there. For features, it works across the whole stack — iOS, Android and backend — from the one-line ask in the tally. It reads the row, finds the code, implements it on both platforms, and writes the release-note bullet.
For support, it writes the reply. Every ticket that needs an answer gets a draft in a drafts/ folder in the repo, dated and named after the person: 2026-09-09-monalisha-add-category.md. Twenty-eight of those this month. The draft knows things a generic assistant would not: that the feature the person is asking about shipped in 2.09 last week, which help article to link, and — through a read-only, sanitised view of the user database — that if someone writes in from a different email than their account, an empty lookup means “wrong address,” not “dead account.” That knowledge is the context layer, and it is the whole reason the drafts are usable.
Then I read it. Every one.
And one number I keep coming back to: 3%. That is how much our users cared about the thing we were leading with. We only found out because an agent counted.
Customer research is not a phase you do before building. It is a feed. Every email is a vote. Every review is a segment telling you who they are. The mistake is not ignoring them — nobody ignores them. The mistake is reading them one at a time and letting the pattern slip through.
We did not build a support bot. We built a system that reads everything, counts it, and hands the count to the thing that can build. The human is still in the loop at the only point that matters: the send button.
If you want this running for your own product — Hermes or a Grok bot on your inbox, Claude Code or ChatGPT in your repo, the context layer and the tally and drafts convention, hardened so nothing leaks and nothing auto-sends — we do it as a fixed-price AI Employee Setup. Small businesses with one inbox start at $1,000; startups with reviews and a repo get the full pipeline from $10,000. Here is what it costs and why. Originally published on Medium.
No. It classifies, tallies, flags, and drafts. Every reply and every review response is read by a person before it is sent. In Pocket Clear’s setup about 80% of drafts go out as written and 20% are edited.
Hermes Agent reads the support inbox and App Store / Google Play review feeds into a tally and writes a weekly research brief. Claude Code works in the repository to implement features from tally rows and write reply drafts. Both are open, self-hosted setups you control.
The first time it took a weekend: an agent with an email integration, a repo the coding agent can work in, and a convention for briefs and drafts. As a done-for-you package, a single-inbox Starter goes live in 2–3 days and the full pipeline in about two weeks.