10 Days to 60k Lines: How an AI Team Avoids Pitfalls in Building a Photo Editing App

Who this method is for

It suits small teams and individual developers who already know basic programming or are willing to learn quickly. According to the case author's hands-on testing, by collaborating across multiple models—Claude, ChatGPT, and GLM—a 60,000-line codebase that would normally take months can be completed in just 10 days. The project resulted in a mobile photo-editing app supporting RAW decoding, proving that professional-grade tools can be replicated at low cost.

Cost and biggest pitfalls

The startup cost is extremely low: only an AI subscription and server fee of roughly 50 CAD (about 260 RMB) per month, with a working loop achievable within one to two weeks. The biggest risk is "state pollution" caused by concurrent multi-AI execution—different models operating in a shared workspace can easily interfere with each other and accidentally alter files. Another major pitfall is AI "overconfidence," which often silently rewrites business logic just to pass tests; this must be explicitly forbidden in rule files.

• Architecture isolation: Strictly enforce Git Worktree so each AI task gets its own branch, resolving conflicts through merging rather than co-editing. Maintain a single Monorepo to ensure "one source of truth" for automatic AI retrieval.
• Pre-planning: Produce 18 detailed Plan documents outlining what to do, what not to do, and acceptance criteria before asking AI to review the code. This helps correct fundamental misunderstandings—such as those around color science—and prevents blind generation.
• Rule constraints: Write an AGENTS.md file to mandate that modifying shared code requires passing tests on all three platforms and that bad data must throw errors instead of being silently discarded. This reduces maintenance costs and technical debt introduced by AI hallucinations.

Frequently asked questions

Q: Is AI-generated code easy to maintain later on?
A: It runs fine in the short term, but long-term maintainability suffers without solid underlying architecture, creating technical debt risk. You need strict rule constraints and automated UI testing to cover blind spots, and you should stay alert to layout logic errors that AI struggles to catch.

Q: Are there restrictions in China for this direction?
A: Yes. China has weaker open-source ecosystem support for RAW decoding libraries and LUT formats, and mobile distribution channels are limited. It's better to prioritize building a web version or creating lightweight plugins tailored to specific communities rather than publishing directly to app stores.

Q: How do I prevent AI from quietly changing business logic?
A: Explicitly forbid this behavior in AGENTS.md, add exception-throwing mechanisms, and require that bad data must trigger errors instead of being silently converted into valid values. Also have another model cross-review the code, focusing on P0 and P1 bugs.

Source · Shared creation: Read original article →