An important point.
I started using AI, going through OpenAI, Gemini, and finally Claude, to ask complex questions as you describe. Over time my use has evolved, with Claude manipulating DEVONthink-hosted data to structure it better via the MCP server and performing queries across data hosted in DEVONthink, Calibre, ArchiMate, Aeon Timeline, email, and my filesystem.
One thing you might try is asking your preferred AI to look at the structure of data and tags in DEVONthink via MCP and asking it for suggestions on how you might more efficiently structure your data or your data processes to make data entry, manipulation, and queries more efficient. Let the AI tell you how the AI can help you.
1 Like
I agree I would not suggest that AI write your novels.
Do you ever have friend or family or an editor give you feedback on your work in progress? Have you thought of using AI to do that - not to write for you but rather to give big-picture guidance and input.
The New York Times recently carried an article claiming that there is evidence to support the point made by @MsLogica: that interacting with the files is an important part of the creative process. A large study of student essays apparently showed that using AI to brainstorm resulted in writing that was superficially more attractive but which was less creative.
Is this the article? If so then it seems to represent one person’s preferred workflow. That may well be accurate for that author, but that does not mean that this will generalize to most people.
Most likely some people would create better work with traditional methods and others would create better work bouncing ideas off of AI.
No. The article I had in mind is https://www.nytimes.com/2026/05/27/opinion/writing-creativity-ai.html. The evidence is said to involve 370,000 students, not just a single person’s work.
Fortunately, my wife is an excellent editor and discussion partner when it comes to books. So, no, I don’t have to use AI as long as I have WI. 
I find brainstorming with an AI useful in the sense that afterward you have a list of ideas to avoid if you want to be original.
2 Likes
I actually use Claude (Haiku is sufficient as this is a low-level task) to proofread the version histories I write, looking for typos, doubled words (e.g., the the scanner), and to check for too much passive voice and make suggestions as needed.
1 Like
A tool that compares my work to the average of its dataset sounds like a terrible choice for anything that would benefit from a personal voice.
1 Like
My personal rule-of-thumb is to use AI when I need (a) repetitive work (b) done to an average quality. The “average quality” is key: AI is great for making a common soup boiled down from countless examples, but is less good at anything that needs originality or nuance.
One great use is for coding parts of my website. AI can crank out in seconds chunks of common code that would take much longer. But whenever the website needs something more specific, it falls down again and again.
I also use AI to proofread my writing for typos, grammatical issues, and word suggestions. Simple useful stuff. But whenever I try to use it for anything that needs be original (sound like me) or creative (in subject, language, or ideas), it stumbles and produces… well, average.
2 Likes
I still don’t really have a use for it as I believe interacting with my files is part of the thinking process, but one thing I have found it useful for twice now is finding a specific stat in a report. […] Both times, I asked the LLM to find the stat I needed and the page number and it worked.
This is the only thing I’ve found LLM technology helpful for: search. I’ve got a utility that I can dump (say) 80 RFCs into, ask it a question about IPv6, and it’ll pull out the info with a citation link to the paragraph of the appropriate RFC where it got the information. I can click to go read it, and it’ll open the document and highlight the paragraph. It uses on-device Apple Intelligence, because I don’t want my money going to any of the cloud AI providers, and I also want to be able to use it with personal documents.
Corrections welcome, but I don’t believe DEVONthink can currently do this?
That is not what AI does. Fronter AI models in particular can analyze and communicate at any level you wish from preschooler to post-doctorate. Yes there are red flags to be aware of - but AI can do much more than “compare to average.”
Understanding you want to use local AI only, you could do this with Devonthink and a local AI model if the size of your sources does not exceed the context window of the LLM.
You could do it with a much larger dataset if you intall a local LLM agent harness such as Goose and then connect that to Devonthink using the Devonthink MCP server.
Understanding you want to use local AI only, you could do this with Devonthink and a local AI model if the size of your sources does not exceed the context window of the LLM.
Right now I’ve got 4.1 MB of RFCs, and I expect to collect more. I also use the software to search 100 MB PDFs.
I tried with a local model, but it was incredibly slow and the results were useless.
[Edit] Also, the solution I’m using takes up 3.3 MB on disk and a few hundred MB of RAM.
It depends what LLM model and what hardware you are running on.
Moreover you said you don’t want to pay for Cloud AI. Not sure why that is but yes that is a notable constraint.
A $20/month Claude Pro subscription running Claude Cowork would work great for your goals.
A $20/month Claude Pro subscription running Claude Cowork would work great for your goals.
Except for the part where I’m absolutely not sending any of my data to the AI companies, even if they were giving away the service for free. (Which at $20 a month they practically are, given how much it actually costs them.)
Anyway, if you have suggestions for specific local AI models that work well, I might give that another try. The one I tried was Qwen3, which articles I saw suggested was good at search + RAG.
Why not?
The one I tried was Qwen3, which articles I saw suggested was good at search + RAG.
What CPU is in your Mac? How much RAM?
Which Qwen 3 version did you run i.e. how many parameters?
“average preschooler,” “average high school student,” “average postdoc”…
Or you could tweak the parameters to compare to a particular writer or group of writers, or a particular type of writing (technical papers, popular non-fiction, etc.).
shrug
It’s still trying to fit my writing to a pre-existing model. Which, as I said, is exactly what I don’t want.
2 Likes
There is no reason why AI needs to fit your writing to a pre-existing model.
What specifically would you want AI to do?