back

under the hood · 2026-07-30 · 2 min

Under the Hood: The Patch Archive

Rebuilding a military patch archive with an agent pipeline: research, redraws, and a render script that turns markdown records into a browsable site.

Grid of redrawn military unit patches from the archive project

I collect the stories behind military patches — soldier-made art, worn as identity. I wanted a home for them that treated the work the way an archive should. Here's how the build actually went, because the process is the part people ask about.

The brief

I didn't ask for a website. I asked for a collection with standards: every patch entry needed provenance, sources, and era context, and the design had to feel like a gallery, not a database. Dark, minimal, mobile-first — you should be able to browse it from a couch.

The AI's first move wasn't design. It was structure: a content schema, a research-sources file, and a research queue — before a single pixel. That ordering matters. When the structure exists first, the design has something true to express.

The iteration that earned my trust

The first pass shipped three seed entries — Merrill's Marauders, the First Special Service Force, VMF-214. Looked great. Then we ran a deep-research verification pass against live institutional sources, and it got humbling:

That last one is the whole method. The archive is smaller than it could be and right about what it keeps.

What stuck as process

Two standing rules came out of it, written into the project so future sessions inherit them: current-designation claims are the highest-rot facts in this domain — live-check them at verification time, and artwork ships only with public-domain provenance recorded per entry (Commons CC0, Army Quartermaster Museum scans, official USMC vectors — each documented).

Same pattern as everything else in my OS: taste and the ask are mine, the machine drafts fast, and a verification gate decides what survives. The iteration loop isn't "make it prettier" — it's "prove it or cut it."