Let’s say I want to write a book. I create a new group and call it My Book. In that group I drop some PDF files containing research I’ve done. I also create a markdown document called Style Guide that has instructions for how to write the book. I also create a markdown document called Outline that has an outline of each chapter of the book. I finally create a blank markdown document titled Draft. How can I have the AI write me a chapter of the book in Draft, following the style guide, with the content that the outline lays out for the first chapter, using the research in those PDFs?
I feel like I’m missing something critical in terms of how to get the AI to interact with multiple documents at once. If the above is possible then this is super powerful and can replace things like Claude Projects. So far though, I’ve tried and it’s hard to get it to actually reference the various documents in the group.
True, but to be clear, I’m not trying to recommend you guys replace Claude Projects. I’m speifically wondering if the above scenario is possible? I’m having trouble wrapping my mind around how exactly the AI stuff integrates with existing documents and groups. The above scenario touches on various different ways AI works with documents, so if you can explain how that would work, it will give me a much better grasp on the AIs capabilities and how to use it.
I did. I even asked the AI chat in the built in help for assistance. But for example, if I click on a project group containing two PDFs and a blank md file and ask it to summarize the two PDFs and write that summary to the md file, it can’t seem to find those files. And if I guide it through searching to find them, it fails to write to the md file even though all the permissions are enabled in settings. These are things the help said would work (and the chat walked me through doing), so I’m pretty sure I did something wrong. Having a better understanding of what the capabilities are and how to use it using the above scenario will help me fill in the blanks.
Here’s an example of a test I just ran right now (Using Claude Sonnet 4.5 via OpenRouter). When I say “It’s selected” I had just selected it. The reason I mentioned not doing the annotations field is I had just tried this before and after some working with it, I got it to find the file but it wrote it in the annotations field instead of the content.
I’ve actually been doing a lot of testing on document detection for AI. We make a ton of assumptions about what AI can see and what it will do. We see the documents right in front of us so surely AI sees them too?!?
One of the things development would have to weigh in on is document selection. Given specific instructions, the selected document supersedes others when it comes to updating the text in another document. Unless you’re using a scripted solution, I would recommend telling AI to create the draft chapter document instead of updating an existing one.
You would think so. So far, Claude is at least able to actually see what’s in the group. GPT 5 earlier was having issues with even that part, so at least that helps guide me in the right direction. I just tried the test again without the annotations instruction and it was able to summarize both documents, but again it put the result in the annotations property instead of replacing the content. It seems to be working better though at least!
Edit - it wrote it in the annotations field of the group itself, it never touched the markdown file
So referencing items by item link in the prompt, though again this doesn’t work consistently when updating the content of an unselected document. And yes, I understand this method doesn’t easily scale to many source documents.
I believe the specific scenario you describe (write a book chapter, following the study guide, using the outline and PDFs as a guide) can be done using the basic concept of an “AI Agent” here:
Beyond that you can consistently/reliably use DT AI for all sorts of complex scenarios if you access the AI via Applescript.
This should be also possible via the chat assistant, at least with a powerful model and a sufficient context window. E.g. by selecting the required documents first (assuming there are relatively few of them), then describing each in few words (e.g. by order) and finally telling the AI to do whatever you’d like to do.
For similar scenario’s I actually use Obsidian with the co-pilot plugin. If the obsidian vault is indexed in DVT then there is a reasonable link to both (not perfect - tagging e.g.). The co-pilot lets you select your model(s), use your documents (md), and actually update your files or write into your files as you request. That way, both worlds are connected (DVT en obsidian). Co-pilot has a free version but I use the paid version so I have no idea if the free version can do this as well. Of course … obsidian is some kind of rabbit hole …
