As I build robot skills for my workflows, I’m seeing how much manual work I do because I’ve used very little automation up to now.
I just built a detailed Robot workflow to handle blog post cleanup—archiving files, moving folders from Desktop to Documents, updating a tracking file, and writing a record of what happened. It works, but it feels complex and token-heavy.
I’m starting to think I built this in the wrong layer.
I don’t currently use shortcuts, Hazel or Alfred, so shifting would mean learning them while continuing to build out my Robot Assistant. That doesn’t bother me, but it does slow things down.
If you were building this today, would you keep refining the Robot approach, or rebuild this kind of workflow with Hazel/Shortcuts as the foundation?
I’m now working on an email workflow and seeing the same issue. I think it might benefit from rethinking the system first (maybe something like SaneBox), rather than building more robot logic on top.
Would especially appreciate hearing from anyone who went down one path and later changed course. Thanks.
For black and write A->B->C kind of tasks, use AppleScript, Shortcuts, keyboard maestro, any of those. Quick and easy.
Use an LLM if you need things interpreted, decisions need to made mid process, error handling, or if there will be unexpected input or results, that sort of thing.
Bear in mind the LLM can write all the scripts for you.
I had a fair number of things that I did in Hazel. The problem was that the formats of the s inputs changed constantly and I spent more time fixing Hazel than it saved me. So for now I’ve devolved to renaming and moving files mostly by hand. I am not planning on moving much if any of my on-going file renaming type stuff in the Robot Assistant since the files are all financial or medical and private.
However, for the file renaming of journal files that was being done in the late lamented Zot File within Zotero I am working on the robot assistance to do that. The reason is the input formats also vary a lot so I need the power of the LLM to make it work.
I’ve asked Claude to help me build token-frugal workflows. We chat back-and-forth about what I want to do and whether I should use an app, have Claude produce a script I can run on my own, or use the robot assistant for all or part of the process. If I can articulate my goal, what my resources are (e.g., my token budget, my app library, things I’m willing to do in the terminal, etc.), how often I need to do the task, etc., we usually arrive at a good solution.
Often it can (and will) write scripts for stuff doesn’t require the kind of analysis and interpretation mentioned above.
I try to keep an eye out for opportunities to do that, but don’t always remember. So I put in my CLAUDE.md that it should aim to be token efficient. That helps a bit. And the Robot Self Improvement stuff I mentioned in that post also helps.
I think if you haven’t bothered with Hazel and shortcuts before, then you don’t have the same stuff to unlearn. Shortcuts is great for certain things, but don’t think you’re at a disadvantage because you didn’t start there.
I put everything into the Robot (both the decision-making and the file operations), which is why it feels complex and token-heavy.
The split you’re describing makes sense:
scripts/automation for A→B→C work
Robot for anything variable or interpretive
My workflow clearly has both. So my next step is to separate them. Keep the Robot for identifying/deciding, and push the deterministic parts into scripts (with the robot helping generate those).
That also reassures me I don’t need to stop and go learn tools like Hazel first. But can do so selectively.
The robot self improvement – yes. I’m going to have the robot review the current skill and help me redesign it with this split in mind. And the token-friendly guidance. Now that I’ve built it, the trade-offs are much clearer.
Shortcuts is great for certain things, but don’t think you’re at a disadvantage because you didn’t start there.