Disclaimer
Is this project open source? N/A - workflow write-up
Is this project completely free? Yes
Is this project vibe-coded beyond the author’s ability to comprehend how it works? No
Community Directory: N/A
Hi everyone,
I wanted to share a workflow I’ve been using for working with Claude and my Obsidian vault.
The problem I kept running into was trust. Claude is useful for summarizing notes, finding related files, and drafting cleaner versions of messy notes, but I do not want it directly editing my long-running vault.
So I started using a simple “proposed edits” pattern.
How it works
At the root of my vault I keep a few folders/files like this:
raw/
proposed/
notes/
CLAUDE.md
raw/ is where I put messy inputs: transcripts, copied text, PDFs, meeting notes, etc.
proposed/ is where AI-drafted notes or edits go first.
notes/ is the actual vault content I trust.
CLAUDE.md is a short instruction file with my vault rules: naming conventions, folder preferences, what kind of summaries I like, and what not to touch.
## The basic rule
Claude can help draft, summarize, and suggest.
But nothing generated by Claude goes directly into the real notes folder.
It goes into proposed/ first, then I review it in Obsidian and move it over manually if it is actually useful.
## Example prompt
Look at my recent notes and give me the three things I should care about from this week, with the file paths that support each.
I’ve found that asking for file paths every time makes the output much more useful. It keeps Claude from giving vague summaries and makes it easier to check the source note before I accept anything.
## What this has helped with
- Summarizing recent notes without losing the source trail
- Turning messy raw notes into cleaner Markdown drafts
- Suggesting wikilinks without automatically changing old notes
- Keeping AI-written text visibly separate until reviewed
- Avoiding the feeling that my vault is slowly being rewritten behind my back
## What I am still figuring out
- Best naming convention for proposed files
- Whether accepted AI-assisted notes should get a tag
- How much context to give Claude before it over-reads
- How to handle conflicts when the real note changes after a proposal is drafted
Curious if anyone else is using a similar proposal-first workflow. I like AI help in the vault, but I want Obsidian to remain the source of truth, not a place where generated text silently accumulates.