Preserving Public Information

I built and tested a small proof-of-concept converter while investigating this request (actually ChatGPT-Solaris built it). It processes Safari WebArchives saved from ChatGPT Share pages entirely offline and converts them to Markdown. The source archives are opened read-only, and the converter preserves headings, lists, fenced code blocks, tables, ordinary links, speaker order, and recoverable citation links. The important technical finding is that the Share page contains a structured serialized conversation in its main HTML resource. The converter does not scrape whichever messages happen to be rendered on screen; it reads that embedded conversation structure, selects the visible user messages and completed assistant responses, and verifies the final message against the recorded conversation endpoint. I tested it on five conversations ranging from 4 to 640 visible turns. All five converted successfully with correct user/assistant alternation and matching final endpoints. I would be happy to provide the inspectable source or the packaged macOS drag-and-drop proof of concept if it would be useful to the team. I would not include any private conversation archives or extracted user content.

A tool for collection archival slivers of the web and web archives - anjackson/sliver