Summary
Every conversation with a coding agent such as Claude Code is saved on your machine, and Cole Medin argues that this history is an underused resource for making the agent better over time. This video shows how to turn it into useful insight.
Claude Code stores sessions as JSONL files containing every prompt, tool call, MCP server use and reasoning step, and deletes them after 30 days by default. Medin starts with a simple approach: ask the agent to review past conversations and suggest the top ten improvements for efficiency and reliability, keeping the output bounded so it doesn’t return a pile of minor suggestions.
He then shows a more robust method using Databricks Free Edition (the video is sponsored by Databricks). Transcripts are uploaded to a permanent volume, parsed into structured tables covering sessions, tool calls and agent turns, and queried through the Databricks CLI and its Genie agent, which can also be connected to Claude Code as an MCP server.
The payoff is a token-efficient, repeatable process for spotting common failures and wasteful patterns, and for turning them into better rules and skills. The approach works with other coding agents such as Codex as well.
📺 Source: Cole Medin · Published September 25, 2026
🏷️ Format: Tutorial Demo
