I've been experimenting with a way to preserve the state of long AI-assisted investigations, which I've called PUFH (Pick Up From Here).
As I've been working through SymPy issues and PRs with AI assistance, I've found that ordinary conversation summaries often preserve conclusions but lose the reasoning that led to them: what was tried, what failed, why certain approaches were rejected, and what remains uncertain. With my ChatGPT account I have exceeded the chat limit 4 times on one project and felt like the next chat always needed a little more help to get back on track.
PUFH is a lightweight Markdown checkpoint designed to preserve that understanding rather than the conversation itself. It distinguishes observations, hypotheses, decisions, and unresolved questions, and provides a starting point for another AI session—or another developer—to continue the investigation.
I'm developing this using actual SymPy investigations as examples.
I suspect others are already doing something similar, perhaps with different terminology or techniques. I'd be interested to hear how people are handling this problem and whether a more systematic approach would be useful. I just recently started using GitHub itself as a place to store updates as I work on a project -- that was an idea I read about in a reddit group.
The idea is still evolving, and feedback would be welcome.
[This message was made by ChatGPT, read and edited lightly by me.]
/c
PUFH: Preserve the understanding, not the conversation.
A good checkpoint remembers not only where you are, but why you aren't somewhere else.