What they did
The poster owns a history of the 106th Cavalry Group that its men printed in Germany in 1945, with rosters of almost the whole unit. Their grandfather, a lieutenant in B Troop, is named twice in it, both times in a forest attack in Lorraine in October 1944. They scanned the book with a phone scanning app and gave all 368 pages to Claude, which ordered the story by troop and date, flagged OCR mistakes and caught duplicate names. Claude then pulled out the rosters, matched names in the narrative to men on them, and generated a page per soldier. Claude Code wrote the 3D scenes with three.js (terrain, trees, weather, figures, gunfire) and adjusted the look by taking its own screenshots and comparing. The poster set two rules: nothing invented beyond the book, and every scene labelled by how strongly the book places their grandfather there.
How it went
They published three pieces as Claude artifacts: a two and a half minute interactive scene of the October attack ending on the real medals, a chapter-by-chapter route from Normandy to Austria with the page scans one tap away, and a searchable directory of all 2,655 men. Sound came from 50 public-domain recordings placed in space. The poster calls it unfinished: the figures are rough, and one figure walked through a tree until they added a build check that fails on collisions. A friend told them the original pages and maps should lead, and that is their next step.
Worth knowing
Keeping the agent from filling gaps took an explicit rule; without it a battle scene invites invented detail. In the comments, a person doing similar work with naval records said frontier models got expensive for OCR at this volume, so they run two local models and hand-check only the lines where those disagree.