In my last thread, I asked how people keep project instructions at hand while working on files. The replies pushed back on how much needed arranging in the first place. Keep the note open, know where your material lives, and get on with the work. One reply put it plainly: “A good system should make the work easier, not become another project in itself.”
That gives me a different starting point for thinking about AI. When you’re doing the work yourself, you can refer to a note as needed and judge which parts matter. When you ask an AI tool to do part of it, you have to convey enough of that judgment for the request to make sense.
Suppose you’re comparing two supplier quotes. The PDFs are in Downloads, the budget is in a spreadsheet, and an Apple Notes entry says that delivery time matters more than the lowest price.
Attaching the PDFs is straightforward. But that sentence about delivery could change the recommendation. The spreadsheet might matter too, although only a few rows are relevant.
A short instruction might be enough: “Compare these quotes, but prioritize delivery before the end of the month.” There may be no reason to share the note or the whole spreadsheet. In another task, the source material itself might be necessary.
This is the part I’d like to hear more about: how people make that selection in practice, without turning preparation into more work than the task.
For those using AI with documents on a Mac, how do you decide what to include in a request and what to leave out? An example of the files you chose and anything you explained separately would help.
The question is too broad and vague. If you’re looking for a rubric that defines how to gather information to hand to an AI for analysis, the only universal answer is: “use what’s necessary in your judgement”.
Probing for specific detailed examples isn’t going to be helpful.
I think this may be a case of a solution seeking a problem. What you describe does not require AI, just your thinking. I would not bother using AI for tasks like those; it is like using a sledgehammer to drive a nail.
I may be incorrect in my assessment, but from reading your posts, I think you are making project and task management far more complicated than your examples require.
Here’s what I’ve learned with Opus 5.5: tell the bot what your goal is, point it at the files containing the information you’ve gathered, and let it get to work. The hard part isn’t choosing what information to provide, it’s articulating your goal. That’s where your time needs to go.
Here’s a recent example from my own experience: I’m the sole person who handles the finances of a very small non-profit. I’ve internalized the workflow and the deliverables and can chug through them without giving them much thought. But the workflow is starting to fray around the edges and and, at some point sooner or later, someone else will need to pick up some or all of the workflow.
I created a project and I told Claude Opus 5.5 that I have two goals: 1) tighten up the workflow and 2) prepare a “how to” manual that my successor could use to take it over in the event that I had to hand it over and might not be in a position to provide much transitional hand-holding. (Life comes at you fast.) Since Claude has already helped me with some aspects of this workflow, it did have some insights into what needs to be done, but certainly didn’t have the whole picture. It suggested that I let it review the relevant email account and provide access to the relevant document repository. In a half hour it 1) outlined what needs to be done when, 2) the tools needed to get each step done, 3) the information needed, 4) where there might be some improvements now, 5) how to stage the document repository in Google Drive for sharing and an eventual hand-off, 6) what information needs to be encrypted, 7) and wrote a draft transition manual.
TLDR: If you’ve paid for a good model, you can just let it cook provided you can articulate your goal.
The choice will depend on the task and the person’s judgment. What I’m interested in is how that fits into the work itself. If you’re already looking through files and know which ones matter, I’d like working with AI on those files to be a natural next step, rather than a separate exercise in preparing everything for a chat.
For a straightforward comparison, AI may be unnecessary. I’m more interested in the cases where someone already wants help reviewing documents or drafting something from them. I don’t think that should require a new organizing system. I’d rather see AI fit into the way someone already browses and works with their files, with less preparation rather than more.
Your handover example is closer to the kind of work I had in mind. You had a clear outcome, and Claude helped work out what information was needed without you having to package it all in advance.
I’m interested in how the work continued from there. When you reviewed and revised the manual, did you mostly stay in Claude, or were you also opening the original documents alongside it?
I set up the project itself in the Claude desktop app; let’s call that Claude.ai to keep it distinct from Claude Code. The Claude.ai project is where Claude and I work on tightening the workflow and writing the manual. By default, Claude runs in a virtual computer in Anthropic’s cloud, and that’s where project memory lives. The manual we’re writing lives there as a Claude doc we both can see and edit. Claude.ai also has full access to the repository of Finder Folders where I store all of the documents relevant to my work with the non-profit.
If I want to work with Claude on a particular document in the repository for something outside of the project, I often choose to do that in Claude Code in VS Code. The Explorer panel on the left display’s the repository’s directory tree, the center panel(s) holds the document(s) we’re working on, and Claude Code is in the rightmost panel. Note that there’s a Claude.md document at the root of the directory that gives Claude Code some context about what’s in the directory and some “rules of the road.”
Unless! I’m working on an Excel spreadsheet or Word document—Claude Code can’t work with those formats natively in VS Code. In that case, I would use Claude in Excel or Claude in Word.