ChatGPT exports contain rich structured content beyond plain text that we currently ignore: tool/function calls and results (tool role messages), code interpreter executions, DALL-E image generations, file uploads, web browsing results, and per-message metadata (timestamps, token counts, model identifiers per message). Claude exports similarly contain tool use blocks and artifact content.
Extracting this into transcript metadata would enable richer episode summaries (e.g., "user asked the model to generate an image of X", "model ran Python code to analyze Y") and better provenance for tool-heavy conversations.
ChatGPT exports contain rich structured content beyond plain text that we currently ignore: tool/function calls and results (
toolrole messages), code interpreter executions, DALL-E image generations, file uploads, web browsing results, and per-message metadata (timestamps, token counts, model identifiers per message). Claude exports similarly contain tool use blocks and artifact content.Extracting this into transcript metadata would enable richer episode summaries (e.g., "user asked the model to generate an image of X", "model ran Python code to analyze Y") and better provenance for tool-heavy conversations.