Token Limits, Wikis, and How to Query Your Knowledge Base Efficiently

Transcript

Hitting Token Limits Faster Than Ever

Good day. My name is Mike from Lone Wolf Unleashed and today we’re going to be talking about what to do with token limits in our AI tools and some ways to capture information and efficiently query repositories. This is relevant for the solo founder because when we go to get the information out of our heads, it has to go somewhere — and when we need to query our knowledge base, we want to do that in a very cost-effective way.

I’m going to take you through that, and that includes Karpathy’s wiki methodology which got a lot of attention this week. Strap yourself in.

This past week I’ve hit my Claude limits faster than I ever have. There has been some chatter about how Opus particularly has been degrading, and some challenges around how many tokens are being consumed to do the same tasks as before. I have certainly found that, yes, I’ve been working on some things that consume more tokens — but my normal activity hasn’t changed that much, and I’m going through tokens at a higher rate.

Because I was already on a Max plan, I was spoiling myself — keeping myself in Opus a lot because a lot of what I do is analysis work. But I’ve started to break down different tasks across different models. Part of why is because I’ve started to hit these limits and I don’t really want to spend more money every month just to maintain my AI use.

Karpathy’s Wiki Methodology

Something that came out this week was from Andrej Karpathy — the same guy who coined the term vibe coding. He’s been working on how to effectively query large information sets with very specific information and have the AI retrieve that in a very efficient way.

What he’s done is create a wiki-type setup in Obsidian. The tweets are out there for you to refer to, and I’m going to have a resource on my website to help you install this in your own ecosystem.

Here’s basically how it works. When you have a knowledge source — I do this a lot with YouTube now — I’ll look at a video, pull down the transcript, and it has a lot of the information there along with the embed so I can watch it again. I can pull this via the Obsidian web clipping browser extension, which inserts a new note with that content into my wiki inbox.

I can then use a command in Claude called wiki-ingest. It ingests the article, creates a summary, and creates stubs — references to other files of similar topics that might be referenced throughout. Then it’s there for me to use.

There’s another command called wiki-lint which goes through and sees what stubs have a lot of incoming references, so you can build up your knowledge ecosystem over time.

Why This Is So Powerful

A lot of people and organisations traditionally build up really big procedures — all their information and processes in a whole bunch of files that have to be managed.

What this wiki approach does is let the AI go in, organise all your information, and build those connections between different places regardless of where in your ecosystem they’re stored — automatically. We don’t have to sit there and figure out which text needs to link to which part of the ecosystem.

It creates a space for us to query the AI against the knowledge base. It’ll look at the index, see if there’s anything relevant there, look at the summary pages, and then if it needs to go deeper it hits the main pages for more information.

This is about 80 to 85% more token-efficient than before. Why? Because it doesn’t have to read the entire markdown page every time it retrieves information. You’re creating a filter — a reverse funnel — for it to go down. Query a little bit, get some information and direction, query the next layer, get more information, go deeper. That’s excellent.

We can take our resources, our procedures, our information — ingest them, codify them, store them in different domain files — all set up to be easily stored and queried.

The other joy of doing this in Obsidian is I can look at all the different stubs, all the connections between notes, pull up my graph view and see how different topics, sources, and articles are all related to each other.

Previously — say five years ago when I was in an enterprise role — a lot of manual work from a team of analysts would go into a process management system. All those connections had to be made by hand. To get visibility you’d have to drill down, find the right file in the right place. You don’t have to do that anymore. I can do this solo for an entire organisation. I don’t need a team of analysts and I don’t need to manually make all those connections.

Being able to visualise how things within your business are connected literally helps you connect the dots.

Target Operating Models and Paperclip

This leads me into the next thing I’ve been working on. My work has accelerated a ridiculous amount recently, and this is how we actually form up target operating models.

If we zoom out of your business and look at what I call my 5Ps framework — the first P is the Profile. The Profile is a very high-level view of your business, and what you end up with is a target operating model.

The target operating model paints a picture of what your organisation looks like — the key players, the key users, the key processes, the systems — all on a page. I’ve just done this for Lone Wolf Unleashed, for the social components and the community and program parts of it.

What I’m now able to do is codify that target operating model, insert it into an agent team, and plan out how to automate as much of that end-to-end as possible.

If I take that target operating model and create an architecture spec, I can feed this into my wiki and know exactly where I’m up to. I can see how decisions made in the past affect what I might do in the future. I can see how I’m going to break down the plan to make a particular future state become true. It’s amazing.

What I’ve started building is a tool called Paperclip. Paperclip came out last month and I’m going to be talking about it a lot more in future episodes — I think it’s incredibly powerful for how solo businesses can be set up to deliver a massive amount of value.

I’ve fed it the target operating model description and architecture, and set up agent roles to help build it out. What we’re going to be able to do is have agents go out to APIs — take a podcast episode, automatically clip it, automatically generate the descriptions and metadata. That’s an extraordinary amount of work taken off the table.

What we’ll end up with is a state where the gap between production of content and distribution is dramatically closed. The time it takes to manage that middle piece — nearly gone.

What does that mean? As a solo operator I can have more meaningful conversations with prospects. I can focus more on community building rather than delivery, content, and marketing. That’s the result I want to move towards — because it will mean I don’t have to keep hiring people to do that type of work. Their time can be freed up for other things and so can mine.

The BA Problem and What I’m Building

All of this comes back to the wiki space — which can be used for any big piece of work, any persistent knowledge. If I’ve got a spec or a build I’m working through, a strategy, or procedures I need to refer to, I can see how all of those things connect and whether I’m on track.

Something this has led me to think about: participants in my program have said, Mike, you’re really good at translating between what the business needs and how to implement that in a platform. I see this with a lot of clients — there’s often a mismatch between the languages both sides of that equation use. I sit in the middle as a translator.

Even a lot of big businesses say they don’t really have the budget for a business analyst. And I can see their systems — they definitely haven’t invested in business analysis. The platforms don’t talk to each other, their users are unhappy, their technology department isn’t delivering well — because they don’t have the information they need. Everyone in that scenario has been set up to fail.

So I’ve been asking: how do I make this better? How do I create a digital AI business analyst assistant that a subject matter expert can sit alongside — one that can prompt, create requirements documents, do workshop plans with the right questions to ask? All with an inbuilt wiki that tracks requirements, constraints, open questions, and outstanding items — all in there.

This is why it’s so powerful. Under any application, within the knowledge ecosystem as a knowledge worker, you can create an agent team that utilises a wiki automatically to deliver more results and tell you exactly where things are up to.

What I’d Like You to Do This Week

There’s been a lot covered in the last fifteen minutes. Here’s what I’d like you to take away.

Go away this week and think about some processes you’re using your knowledge for, and how AI might be able to speed up the outcomes of what that looks like.

What we’re really trying to do is level up our use of AI to manage administratively burdensome tasks so we can focus on the things that actually matter. What are those things? Relationships, client conversations, asking good questions, networking. And outside the business — family time, time with your partner, time with friends, more time exercising. The really meaningful stuff that you know you want to do more of but you’re working too much to do right now.

This is possible. We are in a space now where this is possible.

AI models are getting more expensive to use. We need to be token-efficient. We need to start leveraging this technology to free up our time. Being able to do this means we can remain as solo businesses — we don’t need to keep going out and finding more and more people to do laborious admin tasks.

There will be resources on the website about the wiki setup. Head over to lonewolfunleashed.com/resources and check that out.

Thank you so much for joining me today. I really appreciate it. You could have been doing a million things but you decided to hang out with me and figure out how to use AI more efficiently — alongside your knowledge resources and your wikis. Thank you so much and I’ll see you next week.