How we build
Kynoss Studios is one developer and Claude. I decide what to build and what ships. Claude does nearly everything in between, from the first line of code to the App Store upload.
- apps on the App Store
- 15
- ideas killed: 105 before a line of code, 26 after we'd started building
- 131
- App Store updates since March 2026
- 190+
- how much more Claude reads than it writes
- 800×
From idea to App Store
Every app goes through the same stages, and it can be killed at any of them. Most are.
An idea gets argued with
Ideas start as one-liners. Claude researches each one: who already makes it, whether anyone pays for it, whether it can run on the phone instead of a server. It scores the idea out of 30 and then writes the strongest case against it. I read both and decide. This is where most ideas die.
The riskiest question goes first
Before any real code, Claude tests the thing most likely to sink the app, like whether the phone's built-in AI is accurate enough for the job, and writes down what the app is for. I sign off at each gate, and apps have been killed at the very first one.
It's built in short sprints
Claude builds one sprint at a time and can't call anything done until it has run the app on a simulator and looked at the screen. “The tests pass” isn't proof that it works.
A separate check asks whether it feels right
Three independent AI reviewers judge the experience, not the code: would a stranger understand this screen in three seconds? Would you show it to a friend? Nothing ships without a pass.
Words and pictures are checked against the app
App Store screenshots are made from real captures. Before upload, Claude reviews its own code the way a colleague would and checks the privacy policy against what the app actually sends.
Claude uploads. I press Submit.
Claude archives the build, uploads it and fills in the App Store listing. Sending it to Apple for review is my call.
Then it gets maintained
Every morning a script pulls reviews, ratings, crash reports and sales from App Store Connect. A one-star review or a new crash becomes a written work item, and fixes go back through the same checks. Apps that stop earning their keep get retired.
What keeps it honest
An AI that writes most of the code will also happily tell you it worked. These rules exist because of that.
Rules come with receipts
Each rule in the workflow points to the incident that created it, so nobody, human or Claude, can quietly drop one that seems fussy.
A failed check stays failed
When a test misses its target, Claude can't redefine the target or call it close enough. It stops and asks.
Claims follow the shipped app
Prices, privacy statements and support answers are checked against the app as it actually ships, not against what we meant to build.
The bill is measured
A daily report tracks how much Claude reads. Most of it is Claude rereading its own notes, and a session that rereads too much gets flagged.
The stories behind these rules are in the field guide: what goes wrong when you build with an AI coding agent, and what we do about each one.
What stays with me
Deciding what's worth building. Spending money. Sending an app to Apple. Anything that goes out under the studio's name gets my yes first.
What runs it
A small orchestrator we built on Claude starts the work sessions, hands each one its next item and sends me a morning brief on my phone. The apps themselves don't need any of it: they run on your iPhone, not on our servers.