Your Past AI Conversations May Be More Valuable Than You Think

Your past AI sessions are more valuable than you think. Over the last few months building a biotech tracking app and newsletter, I’ve accumulated hundreds of sessions in coding agents like Claude Code. I recently started treating those transcripts not as disposable history, but as a detailed record of how I actually work - my preferences, my questions, and the decisions I make along the way.

For my weekly newsletter, I had an agent analyze past editorial sessions to understand what I care about: which developments I flag, how I frame them, where I want editorial control vs. where data compilation is the bottleneck. That became a structured workflow where the agent brings the data and I bring the judgment. It took iterations to shape. You try it on, adjust where you stay in the loop, even capture ongoing decisions in a form of editorial memory that informs future sessions. When it clicks, it’s genuinely one of the best human-AI collaborations I’ve experienced.

I applied a similar idea to my clinical trials database. The agent found and read through all sessions where I’d manually caught data quality issues, characterized the types of errors I tend to find, and together we built an automated audit that now runs nightly. Hundreds of one-off fixes across months of work became a set of repeatable checks and follow-ups, one step closer to the ideal of a self-healing database.

None of this is a free lunch - it works because I’d done the real work across many sessions first. But if you’ve been using coding agents regularly, you probably already have that history, and it captures more about how you work than you might realize. These techniques are evolving quickly, and I suspect agents will start surfacing these patterns on their own before long. In the meantime: if you could codify more of your workflow, which parts would you actually want to keep for yourself?