AX notes
AX notes record the rules SugarMount settled on, and the numbers it actually measured, while building and shipping Android apps with an AI coding agent.
This is AX (AI transformation) applied to one-person app development. In the 20 days from the first commit on September 20, 2026, four Android apps (Compress Image, Flip & Rotate Image, Convert Image and HWP Viewer SugarDoc) went live on Google Play.
Every number comes from the git history, changelog and on-device QA reports of SugarMount's app repository. What we could not verify, we say so.
Find a note
Matches titles, summaries, topics and note types.
Browse by topic
Working with agents
How work is split between an AI coding agent and a person, how rules reach the agent, and what we changed after several sessions collided in one repository.
Testing and release
Unit tests, release gates and on-device checks that AI-written code goes through before each release, with the real defects those checks caught along the way.
Android development
Gradle builds, library upgrades, UI and performance problems we ran into while building Android apps with a coding agent, and how each one was solved.
Game development
Generating puzzle game levels with an agent, proving every level can be solved, and tuning difficulty: the rules we settled on and the numbers we measured.
Play Store and monetization
Google Play listings and review, store screenshots and descriptions, and ad setup: the work of shipping and running apps, shared with an AI coding agent.
Website and search
The page structure, Google and Naver search optimization, structured data and design system of this site, built with an agent and explained from the actual code.
Retrospectives and numbers
Looking back over a period of work: what went fast and what went wrong, with numbers we counted ourselves from the git history and QA reports.
Start here
- How we shipped four Android apps in 20 days with an AI coding agent
The whole run, with the numbers. - How to write AGENTS.md: only the rules an AI coding agent gets wrong
How rules reach the agent. - What to check before shipping an AI-built Android app: the release gate and an R8 crash
The checks AI-written code passes before it ships.
Latest notes
Pre-gzipping Android assets, measured: zero APK savings, and .pack.gz became .pack with a FileNotFoundException
Five pre-gzipped Android asset packs were 18 bytes each larger than the APK's own deflate, and the .gz names vanished, causing FileNotFoundException on launch.
A market:// link opens a store chooser on SKT-sold Samsung phones: set the Play package first
On two SKT-sold Samsung phones, three store apps answer a market:// link and Android shows a chooser. Setting the Play package first opened Play directly.
A splash that never moved on: addOnCompleteListener(this) and onStop
Our spelling app used addOnCompleteListener(this) for Remote Config; the listener was dropped at onStop, so a screen-off launch never left the splash.
A reverse-construction generator and the gate behind 1,100 arrow-puzzle levels
All 1,100 Arrow Knots levels come from a reverse-construction generator and pass the G2 pack gate. Fixing the move rule first made each board solvable as built.
Daily users rose after the redesign while organic D1 retention fell from 15.2% to 11.2%
After an August 2025 redesign, D1 retention of our spelling app's organic cohorts fell from 15.2% to 11.2%; 70% of lost daily users were 31-365 days old.
"6.7" < "6.11" is false: comparing Remote Config versions as strings froze content updates
A Korean spelling app compared Remote Config content versions as Kotlin strings, so from 6.10 on, devices stored at 6.2 to 6.9 got no quiz or notice updates.
By the numbers
| First commit → fourth app live | 2026-09-20 → 2026-10-10, 20 days |
|---|---|
| Commits in that period | 59, of which 52 (88%) are marked as co-authored with the AI agent |
| Code | 439 Kotlin files (164 of them tests), about 64,000 lines (as of 2026-10-18) |
| Unit tests | Whole repo: 410 (2026-10-04, three apps) → 1,091 (2026-10-07, four apps incl. SugarDoc), 0 failures |
| Agent instruction file | AGENTS.md, 96 lines, revised 19 times since it was created |
How these notes are written
- Only our own apps. Every number, screen and script comes from SugarMount apps.
- Numbers carry a source and a date. Numbers change over time, so each one says when it was measured.
- Failures are included. A defect found right after release, and its cause, is often more useful than a rule.
Frequently asked questions
What is AX?
AX stands for AI transformation: changing how work is done so that AI takes part in it. Here it means a way of developing where an AI coding agent writes the code, tests and release scripts, and a person makes the decisions and does the verification.
Which AI tool do you use?
Anthropic's Claude Code. Rules reach the agent through instruction files at the repository root (AGENTS.md, CLAUDE.md), and large tasks are split across several agents.
Do you ship AI-written code as is?
No. Before every upload a build goes through a release gate, a launch check of the minified release build, and on-device end-to-end tests. See the pre-release checks note.
Where do the numbers come from?
From the git history, CHANGELOG and QA reports of SugarMount's app repository (sugar-utils, private). Each note gives the date its numbers were taken.