What a failed Space Jam ’96 remake says about LLMs and legacy web

A developer’s attempt to recreate the 1996 Space Jam website with Claude didn’t land-and the miss is instructive. What’s notable here isn’t the nostalgia, but how quickly modern code generators hit a wall when asked for pixel-accurate, era-specific output. Under the hood, late-’90s sites lean on table-based layouts, spacer GIFs, image maps, frames, and tag soup tailored to Netscape/IE quirks. Left to their own devices, LLMs default to semantic HTML, flexbox/grid, and external CSS-better practices, but not period-correct. The result: something “inspired by,” not a faithful clone.

The bigger picture: exact reproduction requires the original assets (tiled backgrounds, GIF sprites, image maps) and deterministic layout behavior that a text model can’t conjure or fetch on its own. Tiny deviations in spacing or asset paths break the look. Worth noting, this is the same friction teams hit when using AI on legacy stacks: without tight constraints, retrieval of archived resources, and tests that enforce time-bound markup rules (think “HTML 3.2 linting”), models optimize for today, not 1996. What’s actually new here is the clarity of the boundary: LLMs are strong co-pilots for scaffolding and refactors, but period-accurate emulation still demands retrieval, tooling, and human curation. Hype aside, “code any website” doesn’t cover fidelity, assets, or browser-era quirks.