So personal curation of knowledge has zero value, because the LLM model will do a better job of it than you because it has access to a wider corpus and is also less inherently less biased?
I don't think so. they're saying documenting your decision process that says how the personal curation happens is valuable but the output of that is not. in Claude language crafting a personal knowledge curation skill is valuable but investing in over engineering the output that can be very cheaply reproduced is not
I guess what I'm not getting about this argument is that the output can be really cheaply reproduced. its a huge pain in the ass to reproduce "understanding" every single claude session (and a waste of tokens). its too much info to put in a claude.md and the biggest win is that it aggregates information specific to your tasks from dozens of different curated sources and then i've curated them to weed out outdated info. fetching all of them over and over is a waste of time and tokens.
the LLM wiki structure I'm using is curated for my projects and my work, gives my agent a single source of truth (therefore not searching endlessly and finding outdated information on confluence, slack, whatever) and operates in a closed system where I can be absolutely sure the information it has about the project is up to date and correct. With a wiki like this for a specific project I can orchestrate entire epic level plans front to back with much much higher implementation accuracy than if every agent in my swarm has to go figure out the correct answer to X question independently. i'm not particularly attached to the specific LLM wiki "shape", but documenting and persisting knowledge for my agents is more important than if i were just coding by myself.
I get that maybe this isn't a marketable service, but my second brain is not "dead weight" - it has eliminated so many real frustration points for me doing ai-assisted programming for a big company where an enormous amount of the text you need to weed through is completely irrelevant or straight up wrong - not to mention the personal knowledge and preferences I have for output being codified somewhere central to the agent.
if agents had real true memory, then maybe "planning is essential, plans are useless". but that simply isn't the case - this is the persisted memory store.
what you're describing is not the typical "second brain" though, it's a persistent decision scaffold memory. I 100% agree and do the same. I have dozens of documented decision structures that are executed as agent skills or workflows and that's a massive time saver. some organizational context files that support the execution of these workflows are also necessary simply because the frequency of use x cost of production doesn't make sense otherwise. now blowing that up to dumping a bunch of one off documents just because you might one day need them, that's the dead weight. I dont think what you're saying is at odds with the point of this discussion
So, a bit like the "Plans are useless but planning is essential" quote... it's not necessarily the output that's important, but what you learn (and in particularly the assumptions you challenge) along the way?
in my experience (and my interpretation of this article) that is the case yes. in your example your ability to do good planning is valuable and generalizable and can help you out of tight corners, but the plan itself as an artifact can become a vanity metric and more of a distraction. a lot of people love selling / buying templates (this is how I organize my schedule every week, for only $50 you can copy my routine) and then that produces zero lift in people's productivity because the artifact sold is decontexutalized and rehydrating in the new person's context, that requires deep thinking and understanding the "why"s which was not sold as part of the package
I never understood pkms because of this. Obsidian felt like a $30 moleskin, when I had an effective stack of $0.69 notebooks. Gmail mcp is nauseating. Still plenty of non personal use cases for these systems.
I think events are a much more valuable datapoint than random bits of textual information in an ever growing wiki jammed full of sentiment and playbooks. Those are nice too, but like “what actually happened” and “what’s in the queue” seem more interesting along with “what’s going on in the world.”
So personal curation of knowledge has zero value, because the LLM model will do a better job of it than you because it has access to a wider corpus and is also less inherently less biased?
Am I interpretting your argument correctly?
I don't think so. they're saying documenting your decision process that says how the personal curation happens is valuable but the output of that is not. in Claude language crafting a personal knowledge curation skill is valuable but investing in over engineering the output that can be very cheaply reproduced is not
I guess what I'm not getting about this argument is that the output can be really cheaply reproduced. its a huge pain in the ass to reproduce "understanding" every single claude session (and a waste of tokens). its too much info to put in a claude.md and the biggest win is that it aggregates information specific to your tasks from dozens of different curated sources and then i've curated them to weed out outdated info. fetching all of them over and over is a waste of time and tokens.
the LLM wiki structure I'm using is curated for my projects and my work, gives my agent a single source of truth (therefore not searching endlessly and finding outdated information on confluence, slack, whatever) and operates in a closed system where I can be absolutely sure the information it has about the project is up to date and correct. With a wiki like this for a specific project I can orchestrate entire epic level plans front to back with much much higher implementation accuracy than if every agent in my swarm has to go figure out the correct answer to X question independently. i'm not particularly attached to the specific LLM wiki "shape", but documenting and persisting knowledge for my agents is more important than if i were just coding by myself.
I get that maybe this isn't a marketable service, but my second brain is not "dead weight" - it has eliminated so many real frustration points for me doing ai-assisted programming for a big company where an enormous amount of the text you need to weed through is completely irrelevant or straight up wrong - not to mention the personal knowledge and preferences I have for output being codified somewhere central to the agent.
if agents had real true memory, then maybe "planning is essential, plans are useless". but that simply isn't the case - this is the persisted memory store.
what you're describing is not the typical "second brain" though, it's a persistent decision scaffold memory. I 100% agree and do the same. I have dozens of documented decision structures that are executed as agent skills or workflows and that's a massive time saver. some organizational context files that support the execution of these workflows are also necessary simply because the frequency of use x cost of production doesn't make sense otherwise. now blowing that up to dumping a bunch of one off documents just because you might one day need them, that's the dead weight. I dont think what you're saying is at odds with the point of this discussion
So, a bit like the "Plans are useless but planning is essential" quote... it's not necessarily the output that's important, but what you learn (and in particularly the assumptions you challenge) along the way?
in my experience (and my interpretation of this article) that is the case yes. in your example your ability to do good planning is valuable and generalizable and can help you out of tight corners, but the plan itself as an artifact can become a vanity metric and more of a distraction. a lot of people love selling / buying templates (this is how I organize my schedule every week, for only $50 you can copy my routine) and then that produces zero lift in people's productivity because the artifact sold is decontexutalized and rehydrating in the new person's context, that requires deep thinking and understanding the "why"s which was not sold as part of the package
I never understood pkms because of this. Obsidian felt like a $30 moleskin, when I had an effective stack of $0.69 notebooks. Gmail mcp is nauseating. Still plenty of non personal use cases for these systems.
I think events are a much more valuable datapoint than random bits of textual information in an ever growing wiki jammed full of sentiment and playbooks. Those are nice too, but like “what actually happened” and “what’s in the queue” seem more interesting along with “what’s going on in the world.”
Your writing is freaking awesome. But...you know this. I ain't gotta keep screaming this on every platform. IYKYK.
my PKM is doc files, drawings, and photos in file folders. and i love it.