Great setup. You should check out Wikilink Types (https://github.com/penfieldlabs/obsidian-wikilink-types). It lets you label the relationships between wiki-linked notes (supersedes, contradicts, supports, etc.) using typed syntax right in the link text. The relationships get synced to YAML frontmatter automatically, so Dataview can query them. For your use case, Claude could write links like [[TWM.TO Analysis|Latest earnings @supersedes previous quarter @contradicts bear thesis]] and you’d get a queryable relationship graph across your whole research vault, not just connections but what kind of connections. Pairs well with the lint workflow you described since Claude could flag contradictions by type rather than just finding them.
Thanks for this. A tangential question: I see you use Claude, and I believe in the past you used ChatGPT. Was Claude's performance on research better or what motivated you to switch?
I switched in February to Claude when cowork came out. I think the gap between the two was bigger in February than it’s now but Claude managed to stay at least even with other LLMs
thanks I have used this and suggest using mother duck as well. :) Getting Claude or whatever to code and the supervise and correct that code ruthlessly is the hack in my opinion. Less token use and more leverage- as that is what they are great at.
Keeping raw read-only for Claude means any wrong number in the wiki can be checked against a file you saved yourself. The ingest step where it talks through takeaways before writing anything is where I'd expect most bad summaries to get caught, like the RIN price read across the refiners' calls.
Thank you very much for your manual. Can you please tell me if you are using the pro or the max plan? And if you are always hitting the limits. Thank you very much.
Great setup. You should check out Wikilink Types (https://github.com/penfieldlabs/obsidian-wikilink-types). It lets you label the relationships between wiki-linked notes (supersedes, contradicts, supports, etc.) using typed syntax right in the link text. The relationships get synced to YAML frontmatter automatically, so Dataview can query them. For your use case, Claude could write links like [[TWM.TO Analysis|Latest earnings @supersedes previous quarter @contradicts bear thesis]] and you’d get a queryable relationship graph across your whole research vault, not just connections but what kind of connections. Pairs well with the lint workflow you described since Claude could flag contradictions by type rather than just finding them.
Found this hugely helpful, thanks Sebastian!
Thanks for this. A tangential question: I see you use Claude, and I believe in the past you used ChatGPT. Was Claude's performance on research better or what motivated you to switch?
I switched in February to Claude when cowork came out. I think the gap between the two was bigger in February than it’s now but Claude managed to stay at least even with other LLMs
Excellent guide and help. Thanks Seb
Glad it helped you, Thomas!
thanks I have used this and suggest using mother duck as well. :) Getting Claude or whatever to code and the supervise and correct that code ruthlessly is the hack in my opinion. Less token use and more leverage- as that is what they are great at.
Keeping raw read-only for Claude means any wrong number in the wiki can be checked against a file you saved yourself. The ingest step where it talks through takeaways before writing anything is where I'd expect most bad summaries to get caught, like the RIN price read across the refiners' calls.
Thank you very much for your manual. Can you please tell me if you are using the pro or the max plan? And if you are always hitting the limits. Thank you very much.
I am on the pro plan for 18USD. Most of the time its enough for me.
Anthropic has a new customer :-)
It doesn't skip information when there is too much with Claude Code, right?