From Digital Tsundoku to Working App: One Developer’s Revival Story
Many Malaysians have their own version of “Tsundoku” – not unread books, but folders of unfinished software projects. One developer faced this with a half-baked shim that connects YouTube Music to the OpenSubsonic API, so it could work with favourite apps like Navidrome and Feishin. The basics were working using tools like ytmusicapi and yt-dlp, but the long list of required functions turned the project into a chore and it was abandoned. To revive it, the developer started fresh with an AI coding assistant, Claude Code, after setting up a FastAPI and Pydantic environment and preparing a CLAUDE.md file of strict guidelines. The human handled the vision, initial setup and reviews, while the AI read the OpenSubsonic OpenAPI spec, scaffolded the server and generated unit tests. Through repeated iterations on error logs and connection issues, the pair finally got music streaming cleanly through Feishin, turning a forgotten idea into a usable tool.
AI Coding Assistant as Pair Programmer, Not Just Code Generator
The biggest shift for side project productivity is treating an AI coding assistant as a collaborative pair programmer rather than a magic code printer. In the YouTube Music–OpenSubsonic project, the developer first asked Claude Code to scan the openapi.json file and create FastAPI stubs for all modern JSON routes. The AI handled boilerplate and route scaffolding, but the human refined the plan, enforced modern Pydantic patterns and insisted on professional unit tests. When integration with Feishin failed, the developer ran the app, captured error logs and fed them back to the AI. Claude adjusted responses, ensuring they returned correctly structured but sometimes empty payloads instead of nothing at all. This mirrors new research directions where tools like the proposed CodeProphet framework focus on planning, dependencies and structured tasks, not just individual snippets. For Malaysians, it means your AI can now read old codebases, suggest architecture tweaks and handle tedious wiring, while you stay in charge of design and quality.
Why This Changes the Psychology of Side Projects in Malaysia
AI pair programmers can fundamentally change how Malaysian hobbyist developers and students think about unfinished software projects. In the revived shim-layer app, the painful part was not the clever streaming logic, but the long tail of about 80 OpenSubsonic routes across 15 categories, plus caching, SQLite storage and file clean-up. These were exactly the kinds of tedious tasks that cause projects to stall for months. With an AI coding assistant handling repetitive routes, caching patterns and disk-saving boilerplate, it becomes less scary to revisit an old repo after exams, a job change or Ramadan breaks. You can ask the AI to summarise the codebase, generate a roadmap and even propose missing features, echoing how research systems like CodeProphet turn high-level ideas into structured plans. This reduces fear of technical debt, lowers the barrier for solo developers with limited free time and makes it realistic to resume a project years later without re-learning every file by hand.
Practical Tips: Using AI Coding Assistants Safely and Effectively in Malaysia
For Malaysians hoping to ship more side projects, the first step is choosing accessible tools. Web-based AI coding assistants like Claude Code can run in the browser and read your repo, while local editors typically offer integrations with similar models. To keep things safe, avoid pasting sensitive API keys, personal data or client code into online tools. Instead, redact secrets, use environment variables and, where possible, work in smaller chunks of code. Structure your work so the AI is helpful, not chaotic: start by asking for a project summary, then a task list, then implementation for one task at a time. Provide specs, schemas or OpenAPI files, just like the developer who supplied the OpenSubsonic specification and CLAUDE.md instructions. For students, a lightweight “Claude Code tutorial” session might include: set up your framework, write a short README with goals and constraints, then let the AI propose a plan you review before any code is generated.
Know the Limits: Why Human Review Still Matters
AI pair programmers are powerful, but they are not a replacement for human judgment. The revived streaming project only worked after the developer noticed subtle issues in responses, ran real clients like Feishin and repeatedly tested end-to-end streaming. AI can generate messy or overly complex code, especially across dozens of routes, and may misuse libraries if not guided with clear constraints such as the CLAUDE.md file in this case. There are also broader concerns. Over-reliance can erode your own problem-solving skills, and blindly trusting generated code raises risks around security, performance and maintainability. Licensing and copyright questions remain important when AI suggests patterns that resemble public code. Research efforts like CodeProphet show promise in automating planning, but they are still tested in controlled settings, not full Malaysian production environments. Treat AI outputs as draft suggestions: review them, refactor where needed, write your own tests and always run thorough manual and automated testing before deploying any side project you care about.