Detailed Analysis
A product manager with no engineering background used Anthropic's Claude Code and a companion tool called Claude Cowork to build a fully deployed, interactive 3D globe visualizing 15,352 AI job openings across 41 countries — completing the entire project in roughly one day. The workflow was divided into two distinct phases: a research and data pipeline phase handled by Claude Cowork, which curated 1,802 companies, scraped jobs from Indeed and LinkedIn using Python, geocoded 4,682 office locations, sourced 1,594 company logos, and generated a full product requirements document (PRD); and a build-and-deploy phase handled by Claude Code, which translated that PRD into a working frontend application. The final product runs on CesiumJS with Google Photorealistic 3D Tiles, vanilla JavaScript, Vite, and Supabase — a stack that Claude Code itself proposed after independently researching a visual reference the author provided and determining that the originally specified Three.js would be insufficient to achieve the desired quality.
Several technically notable decisions emerged during the build. Claude Code identified that the visual benchmark the author wanted — Bilawal Sidhu's WORLDVIEW application — was built on CesiumJS and Google's photogrammetric 3D Earth tiles, not Three.js, and pushed back on the original PRD to recommend the superior stack. The geocoding pipeline achieved 100% success across 4,682 locations using Nominatim with caching and retry logic, and a city-clustering algorithm was implemented to prevent visual overlap in dense markets like San Francisco. The entire QA process was conducted via screenshot feedback: the author would drop an image of a visual problem, Claude Code would diagnose it, push a fix, and Vercel's auto-deploy pipeline would surface a live URL for re-evaluation within approximately 30 seconds. The glowing polyline representing office locations alone went through six distinct tuning iterations and a full 3D-cylinder experiment before settling on a final rendering approach — all driven entirely through screenshot-based iteration, never through direct code inspection.
The project illustrates a meaningful shift in how non-technical product leaders can engage with software development. The author describes occupying a "design/PM brain" role throughout — pointing at aesthetic or functional problems in screenshots while Claude Code handled all implementation details including Cesium scene setup, Supabase query construction, Vite configuration, a scope-mode state machine, and a panel race-guard. Coherence across a project that underwent substantial scope drift — from Three.js to Cesium, from a simple visualization to a full "SIGINT Terminal" design system with government-monitoring aesthetics — was maintained through three source-of-truth documents that Claude Code kept updated with dated changelog entries. This structured documentation approach allowed the author to verify at any point that the deployed application matched the stated specifications.
The broader significance of this project lies in what it demonstrates about the compression of the software development lifecycle for prototype-grade products. Tasks that traditionally require coordination between data engineers, frontend developers, DevOps, and designers — job scraping, geocoding, schema design, database migration, API integration, UI construction, and deployment configuration — were completed by a single non-engineer in a single working day. This aligns with patterns emerging across Claude Code use cases documented elsewhere, including a customer support application built in two days instead of two to three weeks, and ETL tooling that condensed a week of engineering work into approximately one hour. The agentic workflow's ability to run background research agents asynchronously while the user maintained foreground focus further points toward an emerging model of human-AI collaboration where the human functions as a product and design director rather than an implementer.
The AI Jobs Globe project also carries substantive value as a data artifact independent of its development story. With 15,352 job listings filtered to post-February 2026, classified into four AI job types — technical, upskill, executive, and AI-native — and mapped across 1,144 companies and 41 countries, the visualization represents a real-time, geographically anchored snapshot of global AI hiring demand. The decision to tier companies by reputation and handle edge cases like Chinese companies absent from Western job boards through manual research reflects a level of methodological rigor that extends beyond a typical demonstration project. As AI job markets continue to evolve rapidly and geographic distribution of AI talent becomes a strategic concern for governments and corporations alike, tools that make that distribution legible and interactive occupy genuine informational utility — and the fact that such a tool can now be produced in a day by a non-engineer has direct implications for the pace at which AI-native analytical products will proliferate.
Read original article →