One Page Per Person: AI Built a Private WeChat Memory Book in 24 Days, and Three Walls Held
AI-assisted solo development compressed a quarterly timeline into 24 days, but the bottleneck moved from execution to judgment: every ten lines of generated code gets reviewed, and the product's entire shape is a compliance ruling WeChat handed down. Western developers leaning on AI for a real shipped product face the same ceiling: review capacity, not code generation.
A phone call about a relative's child exposes the gap: a face remembered, a name gone, ten minutes of scrolling a dormant group chat. 相识簿 is the fix. Not a contacts app, not a CRM, no friend requests, no public profiles. One page per person, a book only its owner can open.
The build runs 24 days from first commit. Front end is a native WeChat Mini Program, 16 pages and 30+ custom components at 13,000+ lines of WXML/WXSS/JS/JSON. Back end is Spring Boot 2.7 at 10,000+ lines with MySQL, Redis, JWT, and images direct to OSS. AI absorbs the tedious volume: lunar calendar conversion with leap months, unified API wrappers, 80rpx buttons, BEM class names. It even argues a feature down, scoring the relationship graph against replaceability and usage frequency, and the long-image export wins.
Three walls stop the AI cold. It copies birthday phrasing onto death anniversaries, so wording gets rebuilt manually as 诞辰 and 忌日 with grayscale avatars. It can't read compliance: a public memorial page would reclassify the app as social-notes, a category individual WeChat accounts cannot open, so privacy rules are mandatory, not aesthetic. During integration it invents endpoints and fields, turning the human into an acceptance tester: review ten lines, run, send back.
AI proved useful by absorbing the tedious volume that grinds solo development to a halt: boilerplate pages, uniform styling, repeated API wrappers, and it never got tired.
Compliance reshaped this product into something more deliberate: privacy isn't a marketing line here, it's the only configuration WeChat allows an individual account to run.
Plaintext phone numbers flagged as known debt show how AI-paced development shifts priorities: feature velocity outruns security hardening, and the boring fix lands on an unwritten roadmap.
With AI hallucinating endpoints and fields during integration, the human role collapses to acceptance testing, which means reading generated code fluently is the real skill constraint, not writing it.
Why not add the address book? Don't waste tokens in the future (facepalm)
The address book is primarily for communication.