I see there is a lot of excitement in this forum to using Claude Cowork. As a software developer I’ve been using Claude Code every day and I’ve now extended it to projects outside of pure software development. But instead of using Cowork, I’m using Claude code. I believe this is safer because Claude is restricted to only touching files in the folder on your Mac where you run the Claude command. It can work with all types of files, it isn’t just restricted to creating source code.
So I know the command line is scary to a lot of people, but you really just need to learn claude and maybe a few other commands to navigate around the file system. Once you start Claude you can open a Finder window in that directory and use the files as you normally would in a GUI. But if you’re taking the plunge on using a tool as powerful as Claude, I believe learning the CLI is a small price to pay.
I think it’s a bad idea to give Claude access to your email. AI is still prone to hallucinations and I don’t want it messing around with my correspondence with the outside world. But if you want the same MCP servers available in Cowork are also available to you inside Claude Code. I’ve been using the Expedia MCP to plan some trips for later this year and it worked well. It created some Excel files comparing prices of flights and hotels.
So instead of giving Claude free range to your hard drive, create a specific directory and copy the files you want Claude to work with there. I create a new directory for each project, but you could just have 1 big directory with everything.
That is very close to what I’ve done. I create temporary or “source” folders. Claude works in those and when the project is finished, I move the files back to their original locations.
I’m coming to Claude from a programming/web dev background, and one interesting thing I’ve found is that a fair bit of what Claude does is write scripts.
Claude is way, way better at Python than I am, so a lot of times I’ll ask Cowork how much of a task is being done by a script, and how much is actually being done by AI. Turns out that a lot of things I want done consistently can be done with some Python and cron jobs.
For the average user, I would wager that learning a bit more of the command line (both Claude and system-level) aspect would probably open up more productiivty for less AI spend.
I always ask Claude if what I need to do can be done with a script and most of the time the answer is “Yes, and would you like me to write one for you?” (Lately, it’s been asking me if I’d like a Shortcut to trigger it. Since I’ve never expressed any interest in one, I have no idea why.) I started my professional life at the DOS prompt, so using a CLI feels almost nostalgic … Anyway, Terminal lives in my dock now. I use Cowork to do things and Code to make things.
I love Cowork, but not enough to let it play on my hard drive unsupervised. Claude has its own external hard drive for its workspace, and all the files there are copies. I do give it access to my email for information-gathering purposes, though. It is strictly forbidden to even suggest writing a draft, much less a message.
OK, it’s not clear to me that it is restricted in the same way Claude Code is. I’ve seen videos where it is organizing all the files in Documents for example.
Depends on whether you give it permission to your Document folder. And the exact same thing applies to Claude Code, it’s not sandboxed to its starting project folder, it can write anywhere on the file system that you give it permission to do so.
I’ve defaulted to just Googling things and letting Gemini have first crack at it. I have been reorganizing and cleaning up my media library. The scripts have been a huge help.
I also tried using OpenAI Codex w/ Sonnet but it wanted to turn the simplest command-line things into complex Python code. No thank you.
The scripts I have run on all of my media folders so far:
ExifTool scripts
fix some issue with png files that was producing errors
Manually update some metadata fields in .rw2 files
Populate all date fields using the content of EXIF:DateTimeOriginal
Same for video files
Trim some text that somehow got appended to around 10,000 files
Combo shell/exiftool script:
Look for all files which have a date in the file name, matching several different patterns using Regex - use that to populate all of the date fields in each file.
I also had a one-time use script which went through multiple folders, found all files with one specific piece of appended text, checked to see if there was a matching file without the appended text, and, if there was, delete the file without the appended text.
This was for around 50 folders with 2,000 files each. The primary cleanup was to put the creation date into the ‘Date Created’ column in Finder.
The part I have enjoyed most is troubleshooting issues. My favorite fix so far was where the appended text contained an apostrophe using some very unique Unicode that Bash/Zsh did not understand and couldn’t escape. In the end I went into finder, edited the file name, copied the apostrophe, then pasted it into my script. Works great!
In the process I’ve learned a lot about find, zmv, regex and matching, as well as exiftool. So AI has made me smarter!
This thread has been really helpful. Thank you all! I didn’t realise that getting Claude to write scripts was a more efficient use of tokens.
Am I right in assuming that Claude could write a script for apps it has no access to itself? For example could I get it to write a script to pull all pdfs out of Mailsteward?
It depends. If the app is scriptable, then yes Claude will probably be able to do somewhere between a meh job and a very good job. If it’s not scriptable, and doesn’t have any other automation surfaces then good luck.
Perhaps run it past Claude, itself? Looks like Mailsteward has no AppleScript or Shortcuts support. However, it is essentially an SQLite database, so perhaps explore with Claude if a Python script referencing the SQL db might do what you are after?
Claude, in my experience, has been great at writing Shell scripts, JavaScript, Google Apps Script, AppleScript, Python, Drafts actions, Terminal commands etc etc. Of course, caveated with me having very limited needs.
I never go beyond the free tier of Claude. I fall into the camp of I “know,” certain things can be achieved between programmes on my Mac, but I’m a lazy programmer/analyst. So for me, this use of an LLM is liberating. It helps me boost the use/productivity – call it what you will – of my existing stuff. Also, if you have an idea of, say, “can I send x to y, and have it processed on a schedule?” Claude is a superb sounding board for exploring the possibilities, and likely writing the code to facilitate. Of course, good/clear prompting is a given here.
With all that said, I’m not in a position/interested in spending big sums (for me) on tokens to get Claude to work on my behalf or use MCPs. I of course recognise there are benefits to this, but on the other hand, as I’ve raised elsewhere on here, I do wonder how much “work” is being offloaded to these energy guzzlers, at not inconsiderable expense, that could just as easily be covered off by existing apps/conveniences on a Mac.
LLMs can feel awe inspiring as an end user. But the associated infrastructure behind the curtain to make that magic happen is vast with myriad associated costs/implications. As was suggested on a recent podcast I listened to, in terms of the journey to “true AI,” they could be described as analogous to alchemy, which was debunked and in turn gave way to chemistry.
My use of Claude falls into 2 categories: write scripts, provide options or research.
For python script writing the biggest issue I’ve run into is giving Claude enough samples and test cases of what I need to get the scripts to work properly. Typically I am doing this with things I do not want to share with the upstream data center so I have to either create or find the edge cases that will force a better script. I know MacSparky says he runs a skill about 5 times with corrections but for me it’s more like about 10 times with sample data before I can reliably use it and even then there is usually some cleanup afterwards. OTOH it’s allowed me to tackle some long term projects that I never had the energy to do by hand.
I now have scripts that can process and manage the round trip connections to Readwise Reader with me changing highlights that Readwise said could not be done for 3 years and that’s just the start. i have scripts that produce my written calendar reports that solved an Apple bug that has been present for the past 4 operating systems.
For options a typical prompt is
Read the CLAUDE.md file. Read the Skills and the Skill Index for relevant context.
IIs it feasible to build a functional EID livestock tag reader (ISO 11784/11785, 134.2 kHz, FDX-B and HDX) from commercially available parts that a person with a soldering iron could assemble, targeting a retail price of $300 USD? What are the core components, their costs, and the critical design constraints? Document all references and sources.
That prompt found some home built EID reader info I was not aware of. It also confirmed some assumptions I had about kits and cost of sourcing the components. There were some other pieces that were more helpful that I was unaware of but overall it presented me with a nice report of options that I can then work from going forward.
In my (admittedly limited) experience, it’s just the opposite. Cowork does not have access to any folder besides its project directory - even if you ask it to access other files, it can’t. As I understand it, it’s basically a little sandboxed VM that can only see what you explicitly add to it. You can’t even ask it to review the contents of other projects or conversations in any straightforward way.
Code has full filesystem access including the ability to install and use new developer tools. It asks you for permission for various operations but can do much more damage and/or data exfiltration.
If I have this wrong, please correct me, but that’s what I’ve seen in my limited experience.
Claude has its own external hard drive for its workspace, and all the files there are copies. I do give it access to my email for information-gathering purposes, though. It is strictly forbidden to even suggest writing a draft, much less a message. 