I’m looking at the sales page for the MacSparky Robot Assistant Field Guide, and I’m on the fence. It might be the idea of there being an Obsidian vault with it that’s making me think it’s a PKM/“second brain”-centric thing. But it talks about calendars and such too.
Has anybody here gone through it? Can you give me an example of a real workflow that you’ve created based on the info in the guide?
I’m also undecided. I’ve bought most of the field guides, but in this instance, I need to see practical examples of how it would benefit my productivity, rather than just being cool before making a purchase decision.
I’m working my way through the guide right now and I think it’s something different from a classic PKM system. It’s about crafting your own agent to handle things like email processing, calendar management, task management, contacts/CRMs, content workflows, and integration with other tools. It’s more about leveraging Claude Cowork to handle what David Sparks refers to as “donkey work” rather than building a system to manage notes and the like. The fact that it’s based on Obsidian has more to do with grounding your personal agent in a managed collection of markdown notes than “linking your thinking” or the PARA method.
The first part of the course—about three hours worth of video tutorials—covers the basics of get your robot up and running step-by-step, e.g., learning how your robot “thinks,” creating skills, connecting other tools, building workflows, managing your robot’s files, and robot maintenance. The second part of the course, the workshops (which will be recorded if you can’t attend) builds on that foundation to tackle the donkey work—email processing, task management, etc.
I purchased the Robot Assistant Field Guide more for what it will teach me about working with agentic AI than for the specific donkeywork topics. I’m pretty sure I can map the skills I’ll learn in the course to my specific areas of concern. I could probably have cobbled together a syllabus to learn this stuff from YouTube, Substack, blogs, or just asking Claude—but I know from experience that David’s courses are thorough, in-depth, and well thought out in terms of learning something step-by-step. I hesitated because of the price, but decided my time is valuable and these are skills I want to learn asap.
Did you watch the Robot Assistant Field Guide Introduction that David put up on YouTube? It ends with a demonstration of the daily brief his robot generates as an example of what you can build.
No doubt this is a lack of imagination, but for the life of me, I cannot think of things I would want an agent to do related to any of those items listed. I have used Cowork extensively to reorganize files, to create a sophisticated system within iA Writer for my book project, and to compile research. As far as handling email, task management, contacts, and calendar, however, I cannot think of how it would do that for me in a way that would be both productive and comfortable (i’m not keen on the idea of having AI compose my emails – I’m sensitive to the idea that if I am corresponding with someone, it should be me that they’re hearing from, not the machine).
That said, I am very open to hearing ideas, as I suspect I am simply not being imaginative enough.
I did. In my experience, consolidating all that information into a single print page would not be beneficial. I already have access to that data through my applications, and printing it would not provide me with any additional value for my specific needs. While it may be helpful for others, I do not believe it would be meaningfully beneficial to me.
I’m fairly certain that such a robot assistant would take a lot of work off my hands and save me time. So a good field guide would indeed be useful. What I have major concerns about, however, is entrusting an AI company with so much personal data. We’re talking about emails, calendars, notes, parts of the file system, etc. As a European, am I being too cautious, or what do you think?
Well, it is just an example of what an agent can do. I don’t think it’s the end goal. I’m not interested in that kind of daily brief either, but I am interested in building robots, which is the reason I signed up. Your Mileage May Vary! A lot of the in-depth material I’ve found on line is aimed at coders; I’ve found it frustrating to translate “here’s how I run a dozen agents in parallel to manage my codebase” to “I need an agent to help me with this important but donkeywork-laden area of my life.”
Haven’t seen the course, but you could other storage options with Cowork, for example Apple Notes instead of Obsidian.
Someone in the forum uses Devonthink with it
@rkaplan uses DevonThink with Cowork Using Claude’s Dispatch you can send work to Cowork with Mobile - #5 by rkaplan
AI agents are immature right now, and the field is constantly changing. I wouldn’t do anything production-critical. Even if it doesn’t result in some catastrophe, anything you set up now is likely to break repeatedly.
I will periodically test things in a playground to stay in touch with what’s possible.
Companies right now are scrambling to figure out how to integrate tools into their products, and things will get messy. (My perspective is health care. Don’t ask me how I know.)
My biggest gripe is that ML/AI is much broader than LLMs — but presently, LLMs have made the most progress; LLMs are the hammer, and everything is a nail. In the future, you will see a much wider variety of ML models, many of which can easily run on-device without relying on large models on cloud servers. There is a big movement toward tiny models in the research space.
Another issue right now is cost and the context window. As interactions with LLMs progress, all of that history gets retained in a context window. This context has to be sent back to an LLM with each subsequent request, whereupon it is parsed again and again into tokens for use by the model. There are strategies to prune and summarize the context, but presently there is no way to escape the fact that this is how LLMs work. You have seen this in action if you have noticed that LLMs get slower the longer you interact with them and that the generated output starts to drift from original intent. And that’s why the number of tokens passed (that you are charged for) goes up.
So, context keeps getting shuttled back and forth. This is a huge I/O issue, and server compute is tremendous. This leads to a lot of electricity usage by these centers… and a lot of cost for customers.
Something that needs to happen is to bring the model to the data rather than sending the data to the model. This already happened in the database industry, where database systems can now do analyses, not just storage and queries. This will happen in ML/AI with edge computing, where models get run on-device, exactly where all the context already lives.
I’ve bought it, and it is not really a PKM thing. But I can’t verify any of the workings (although I am watching the videos) because I can’t get Claude to even get past the first interaction in Lesson 3 to get it set up.
It’s very different from a PKM system.
I’ve been working through it for several weeks now. I use it to give me a daily briefing each morning about events for the upcoming day, tasks, etc., and I also use it for a shutdown routine each evening where I review what I did that day (it reads my Noteplan daily notes and journal entries), and previews the next day (tasks, calendars) and I select my three primary tasks, and writes them to the proper section in my NotePlan daily note. In addition, I have a process where it helps me to edit my dictated blog posts and then publishes it to WordPress. Further, it walks me through my weekly reviews, my monthly reviews, and my quarterly reviews, providing me with all the background information I need to understand what I’ve done in the past time period and what’s coming up.
If you are interested in a robot assistant that is able to look at multiple apps and to actually write things to your apps if you want them to, and to help you with processes that are more along the lines of grunt work, then get off the fence and buy it. 
I have several specific things I would like it to do.
I would like to point it at my subscribed calendars our Synology NAS and have it create skeleton .md files with basic info from the calendar as a start for further work in my “real” obsidian vault for historical documentation.
I would like to build a better index and reference from my Zotero library of Scientific papers or even move them into a specified folder in my Obsidian vault and compile my several research projects into some concise summaries with links to original sources and my highlights and notes.
In a coding task I’d like to have it build a way for me to extract full text email messages from Thunderbird and embed them in a note as well as a link back to the original message in Thunderbird. I often need to see certain emails in the context of other data and want that actually in Obsidian. I do not use IMAP mail I want the effective “hot key” to be only on a single selected email not on a bunch though. Most email will stay on my system in Thunderbird.
I would like it to gather all the various epub files I have into a unified “Books” system. Integrate with my Readwise reader Highlights or even us Claude Code to build a Reader ap that is fine tuned to my specific needs since even several years later Readwise is ignoring what is my most wanted feature, round tripping notes and highlights on what I read from Obsidian back to the reader and so on. I take quick notes as I read, refine them over time in Obsidian, take notes again at different times, want to link into specific highlights and notes across many different documents and then see al the changes back in the doc when I read it again along with a datetime stamp so I can see my progression and what’s important.
Extracting the dates we watched movies or TV shows and cross reference with themes and info from my scientific papers and book datasets.
I am very concerned about that as well. So I can’t use it for critical and frustrating “donkey work” tasks of renaming bank statement files, or doctor reports or medical testing data.
But there are enough other things that I could use it for that I want to learn about it and make a choice on whether to use it or not. I’d much prefer a lighter weight but locally run version of the model if at all possible.
Right now however, my most frustrating part is I can’t even get Claude to do the basic stuff that is in the FG as I keep getting access errors.
I’m setting up my robot to do very different things—prepare a detailed summary of a lengthy email conversation, for instance. David has set up the assembler in such a way that you really can build a robot assistant that does what you need it to.
I’m concerned about this as well. I found this guide re Claude Pro privacy helpful.
Here’s what the guide covers:
This comprehensive guide examines Claude Pro’s data privacy in 2026, including the October 2025 opt-in model, retention periods, encryption standards, GDPR compliance, and how it stacks up against competitors. Whether you’re a freelancer handling client data or a professional exploring AI tools, understanding these privacy protections is essential.
Note: I have no idea what AIonX is or does, but the guide is useful …
You only give it access to the personal data you’re comfortable giving it. It doesn’t have to ability to go snooping around unless you give it permission.
I am only on the 4th (short) module and I like it.
David @MacSparky does a load of heavy lifting behind the scenes, and I was going that anyway, and now I don’t have to.
I won’t use it exactly as he does, but I had used Claude loads already and the guide will save me a lot of time.
He’s really good at this sort of stuff. Makes it look easy!
Thanks @Clarke_Ching. I really struggled with the answer when teaching this material, whether I just give you the skills or teach you how to make the skills, but the old thing about teaching Amanda Fish seemed to be the best answer here.
I thought for a second, “Who is Amanda Fish? Why is David teaching her?” Then I finally got what you were trying to say. 