Work & Co engineer spent two years integrating Copilot, Cursor, Claude, and ChatGPT into production web development. The result is a practical taxonomy of where AI earns its place: codebase archaeology, dependency triage, cross-file refactors, unfamiliar tech stacks, test generation, internal tooling, and legacy code rescue. The GLSL shader case is the sharpest example: a complex animated gradient delivered in two days on a language the developer barely knew.
The responsible developer framing is the article's actual contribution, not just a disclaimer. It defines a working standard: AI output gets verified against official docs before shipping, secrets and PII never enter a prompt without policy approval, and pull requests still meet the same review bar as hand-written code. The plotly.js migration from 2.35.2 to 3.1.0, where axis labels silently disappeared, is a concrete case study in how to query an AI for breaking changes and then confirm the answer against the official migration guide before touching production.
The piece goes deeper than surface prompts. It includes specific prompt structures, like the architecture overview prompt that asks for entrypoints, routing, auth, data layer, build tooling, and five files to read in order, alongside a Kent C. Dodds testing principles workflow passed directly to the agent. If you use any JavaScript framework in production and want a replicable AI workflow with guardrails, read the full article at Smashing Magazine.
[READ ORIGINAL →]