technology
Swarm: one engine for six Hong Kong directories
A recruitment directory and a warehouse directory contain different businesses. Much of the software they need is the same: company records, locations, categories and pages that connect them.
The 2025-26 Swarm build used one Astro codebase for six Hong Kong directory subjects. A SITE_SLUG setting selected the data and site configuration. Headhunter.com.hk was the recruitment example and a place to test the content structure.
Build the relationships before the pages
The data layer used SQLite and JSON. A separate knowledge vault represented companies, districts, stations and industries as linked Markdown notes. Those records supplied the material for the site.
For recruitment, that meant a company could connect to its district, nearby MTR station and service specialisations. A district page could then lead back to the companies in it. The useful unit was the relationship between two records, rather than another page written in isolation.
The research and enrichment work started from public business information. AI helped classify services and compose descriptions from that material. It also exposed the weakness of the approach: a company with little source information still gives you little to say. Longer generated copy doesn’t repair that gap.
We used DataForSEO for keyword research and explored structured pages and an llms.txt endpoint as part of a search-visibility experiment. The question was whether richer records and explicit relationships would make the directory more discoverable.
Reuse the engine, keep the subject specific
The shared engine handled page generation, layouts and deployment to Cloudflare Pages. The directory configuration and records supplied the subject. That let the project cover recruitment, audit, renovation, warehousing, water services and education without maintaining six independent front ends.
Sharing code doesn’t remove the editorial work. Each directory still needs useful categories, sourced records and a reason for one page to link to another. Those choices belong to the subject you’re describing.
Our later keyword-research work came back to the same decision: define what the page should answer before producing it. In Swarm, the templates multiplied the editorial model. That made getting the model right the part worth spending time on.
Interested in this?
Learn about Technology →