QoderCN + Qwen 3.8: One Prompt Yields a Physics Video, a Disassemblable 3D Watch, and More
The cost of turning a half-formed idea into a working, multi-file project drops dramatically when an agent handles the full pipeline, including build verification and layout repairs. Western developers evaluating open-weights models from China now have a concrete benchmark: Qwen 3.8-Flash plus a capable agent orchestrator can produce polished, interactive prototypes in minutes, which shifts the competitive baseline for AI-assisted development tools.
A single agent orchestrator can now take a single prompt and turn it into a full, deployable project. In this field report, QoderCN, paired with Qwen 3.8-Flash, handles four distinct tasks: generating a personal portfolio site and publishing it to a live domain, producing a Memphis-style animated video about Newton's First Law (including TTS voiceover and timing data), building a 29-part interactive 3D watch with disassembly, and creating a WebGL face-pinch prototype for a developer's homepage.
The agent doesn't just write code; it structures projects into multiple files with clear responsibilities, runs automated builds, launches local preview servers, and even catches and fixes layout issues like a transparent canvas blocking a theme toggle. For the video task, it reads a custom Skill (one-minute-remotion) that encodes the entire production pipeline, from scriptwriting to Remotion rendering and MP4 export. The author had to tweak a few phrasing sentences in the script, but the technical pipeline ran unattended.
The pattern across all cases is that QoderCN behaves like a junior engineer with a strong grasp of project structure and tooling, not just a code generator. It splits geometry, parts, controls, and entry points into separate source files, centralizes shared dimensions in a specs module to keep 2D and 3D consistent, and runs `npm run build` before previewing. That engineering discipline is what turns a promising demo into something maintainable.
QoderCN's engineering value lies in project management rather than raw code generation: it structures files, runs builds, launches previews, and self-corrects layout bugs, which is what separates a toy demo from a maintainable project.
The author relies on a hand-crafted Skill file to encode the entire video production workflow; without such structured playbooks, the model would likely flounder, suggesting agent effectiveness still depends on domain-specific templates.
Even with full automation, the author manually adjusted script phrasing, meaning the human still owns content judgment; the agent handles mechanics, not editorial quality.
Hands-on impressions split between satisfaction with QoderCN's near closed-loop implementation once a plan is fixed and complaints about slow, laggy performance. Progress in Chinese-comment understanding and local framework fit is acknowledged, but deep-directory context retrieval is flagged as the next hurdle. The generated teaching animation also earns a nod for what it might have changed in school physics.
Qoder is decent now; once the plan is set, it can almost fully close the loop on its own.
Domestic code models have made obvious progress in understanding Chinese engineering comments and adapting to local frameworks. The next key is still the efficiency of context retrieval in deep project directories.
Slow and laggy again.