DEV Community

Discussion on: Claude Code + Obsidian: Build Karpathy's LLM Wiki From Your ChatGPT History

Collapse
 
alexcodebytes profile image
Oleksandr •

The most valuable part of this approach isn't the Obsidian integration or even the export script. It's treating conversation history as source material rather than as accumulated chat logs.

One thing I'd be careful about is preserving uncertainty and context during ingestion. A discussion that explored an idea isn't necessarily a decision, and a decision made at one point may later have been reversed. If the wiki turns those distinctions into definitive statements, it can quietly become less reliable than the original conversations.

I'd also want a way to trace every important claim back to the exact exchange, not just the chat containing it. The git diff helps review what changed, but provenance is what makes the knowledge trustworthy over time.

Have you considered adding explicit statuses such as proposed, decided, rejected and superseded to the wiki schema?

Collapse
 
ashimanski profile image
Artyom Shimanski •

Good question. You're right that the risk is the wiki sounding more certain than the chats ever were. Half of my ChatGPT history is me thinking out loud, not making decisions.

I hadn't added statuses yet, but it's an easy change: one more rule in CLAUDE.md. Something like proposed / decided / rejected / superseded, plus "a chat that only explores an idea is never 'decided'". Lint can then flag old "proposed" items that never got resolved.

For provenance, I think the simplest fix is to number the turns in the export (## 3 You, ## 4 ChatGPT). Then a claim can link to the exact exchange with [[chat#3 You]] instead of the whole chat. raw/ never gets edited, so those links stay valid.

I'll try both on my vault and update the post.

Collapse
 
alexcodebytes profile image
Oleksandr •

That seems like a solid improvement, especially the rule that exploratory conversations can never automatically become decisions. The distinction between proposed and decided is probably one of the most important safeguards against a wiki gradually turning speculation into established fact.

Numbering individual turns also solves a much more practical problem than linking entire conversations. It gives you a precise trail back to the original reasoning, including the context, objections, and uncertainty that might disappear during summarization.

I particularly like the idea of linting unresolved proposals. A knowledge system shouldn't just accumulate conclusions; it should also expose decisions that were never actually made. That could make the wiki more trustworthy over time rather than simply more polished.

Thread Thread
 
ashimanski profile image
Artyom Shimanski •

Updated the post with both:

  • The CLAUDE.md now has a Status section: proposed, decided, rejected or superseded. A chat that only explores an idea can only give "proposed", and "decided" has to link the turn where I actually said so. Lint lists proposals older than 30 days that nothing resolved.
  • The script numbers the turns (## 3 You), so a claim links the exact exchange: [[2025-10-09 Pricing page copy#3 You]]. The gist is updated too.

Thanks for pushing on this. "Expose decisions that were never actually made" puts it better than I did.

Thread Thread
 
alexcodebytes profile image
Oleksandr •

This is a really meaningful improvement. The distinction between “proposed” and “decided” is especially important — otherwise, an AI can gradually turn a brainstorming conversation into a decision nobody actually made.

I also like the exact turn references. They make the system auditable instead of relying on someone’s interpretation of a long conversation. And the 30-day unresolved proposal check is a nice touch: it prevents ideas from quietly becoming permanent assumptions.

The bigger win is that you're not just documenting decisions anymore; you're making it harder for the system to invent them. That's a much stronger foundation. 👏

Thread Thread
 
ashimanski profile image
Artyom Shimanski •

Thanks, that's a better way to put it: the risk isn't the wiki missing decisions, it's the wiki inventing them. If you try the schema on your own chats, I'd like to hear where it still gets things wrong.