Detailed Analysis
A developer has publicly documented the creation of Noctis, a Chrome extension that routes browser traffic through proxy protocols including VLESS and VMess via a local sing-box instance, while leaving the host machine's network traffic unaffected. The architecture separates concerns cleanly: a TypeScript/React frontend extension handles the user interface, while a Go helper application manages the actual networking through native messaging, writing sing-box configurations and maintaining process stability. The system supports rule-based routing by domain, IP, geoip, and geosite criteria, allowing users to direct specific traffic to different servers or block it outright. The practical use case driving the build was the need to run a personal proxy for desktop applications like Telegram while simultaneously routing browser traffic through a corporate proxy — a problem common to developers operating across personal and professional network contexts.
The developer credits Claude Code as the end-to-end builder of the project, noting explicitly that they are not a strong frontend or Go developer and that the AI workflow was therefore central, not incidental, to the project's completion. Critically, the developer describes a disciplined methodology: features were never implemented before a planning phase was completed. Claude Code's Explore subagents were used to map the existing codebase, a written plan was produced and reviewed by the developer, and only after human editing of that plan was implementation permitted. Thirty-six such plan files are preserved in the repository. This structured approach, combined with a test-first development strategy using Vitest and strict TypeScript compiler settings including noUnusedLocals and noUnusedParameters, served as a quality enforcement layer that prevented the mixed-language codebase from degenerating into unmaintainable complexity.
The developer also details several workflow automations built with Claude Code's assistance. A custom slash command called /release was created to read commit history, draft a changelog, bump version numbers, cut a git tag, and trigger CI publication — reducing a multi-step release process to a single command. Additionally, when a specific mistake occurred (a tag-naming error that corrupted a mirror), the developer instructed Claude Code to save a persistent memory note about that error class, and the system subsequently avoided repeating it. These details illustrate a feedback loop in which the developer treats Claude Code not merely as a code generator but as a stateful collaborator whose behavior can be shaped by explicit constraints and accumulated context.
The account carries broader significance as a concrete example of AI-assisted software development enabling non-expert developers to ship production-quality, architecturally complex projects. The combination of a browser extension, a compiled Go binary, native messaging IPC, and external process management (sing-box) represents a non-trivial integration surface that would historically have required deep expertise in multiple domains. The developer's explicit emphasis on constraining the model before coding, and on building automated verification mechanisms rather than trusting the model's self-reported completion status, reflects an emerging best-practice framework for AI-assisted development: human-defined plans as guardrails, automated tests and type systems as objective truth arbiters, and model judgment subordinated to verifiable outputs.
This case study fits within a broader trend of Claude Code being deployed as a primary development agent rather than an assistive tool, particularly among developers whose domain expertise lies outside the languages or platforms required for a given project. Anthropic's Claude Code product, which supports agentic workflows including subagent spawning, plan-mode exploration, and persistent memory, is increasingly being used to bridge skill gaps on projects that would otherwise require larger teams or longer timelines. The Noctis project demonstrates both the potential and the caveats of this paradigm: the methodology succeeded, but only because the developer imposed disciplined structure on the model's behavior, suggesting that human oversight design — not just model capability — remains a determinative factor in AI-assisted software quality.
Read original article →