How I Keep Up With AI

I've developed a learning system that answers the question "How do I keep up with AI when there's a new breakthrough every week?" I'm quite pleased with it.

My process for building with AI so far has been to identify an outcome I want and discover the necessary components to accomplish the outcome. In this way I've built multi-agent systems with orchestrators, sceptics, verifiers, SKILLs, MCP, RAG, tool capabilities, etc.

Over the past year the rate of innovation has been incredible - a combination both of new capability through creation (model improvements, etc.), but also people combining known techniques to get novel outcomes. Ideas spawn seemingly overnight, fall into best practice, and everyone progresses.

I realised I was hearing about improvements from influencers, email subscriptions, and Twitter feeds but often didn't care for other people's analysis. It was typically oversensational, undertechnical, and designed to farm engagement rather than inform. I needed a way to stay on top, reduce noise, and make the technology immediately available.

I built a tool that does three things.

Step one: I pull from Anthropic's 4 RSS feeds and apply a filter mechanism. My agent is sceptic by default, meaning that it prioritises excluding information from my view rather than including it. This produces a much more refined output. What I'm hunting for is build primitives. I want the building blocks of something excellent. If my agent flags an article about a build primitive that I don't already have in my library it does a web search and enriches its understanding of how it works, how it's applied, and reduces it to an instruction Claude Code can understand.

Examples of primitives that would make their way into my library:

Code execution with MCP: Instead of cramming descriptions of every tool an AI agent might use into its limited working memory, you store those tools as code files the agent can browse and run only when it actually needs them. This keeps the agent's attention free for the task at hand, and lets ordinary code do the heavy lifting of moving and filtering data rather than forcing it all through the AI. In one of Anthropic's examples, that cut the memory used from about 150,000 tokens to 2,000.

Filesystem as agent memory: An AI agent normally forgets everything the moment its short-term memory fills up, so any task longer than that gets lost halfway through. Letting the agent jot its progress and notes into ordinary files means it can pick up exactly where it left off after a reset - the same way a person leaves themselves a note before stepping away from a half-finished job.

When I first built the library I scraped every repo I've built to uncover my own primitives and catalogue my novel creations. I compared my builds against the Anthropic feed, deduplicated where necessary, and now have a good combined library.

Step two is a human in the loop verification process. My tool surfaces the build primitives it has encountered that week, and one by one I decide whether they belong in the library or not - whether they're mentions of existing primitives and don't merit appending additional information, or if there's something genuinely novel that needs to be there.

The result is a clean library of build primitives, each with a description simple enough for me to understand conceptually, and detailed enough for Claude Code to execute on.

Step three, the final step - and the most important - is putting the library to work. Now, when I want to build something new I work with Claude to carefully define the desired outcome of my build - we work through all potential features, capabilities, and limitations. Once that is settled I expose Claude to the primitives library and have it cherry pick the building blocks that would make the desired outcome possible.

The end result is that I'm able to build systems that make the best use of the existing concepts and technology today, and learn how they work in a practical way, as I troubleshoot, refine, etc. I'm very pleased with the workflow - I'm reducing token consumption, caretaking cleaner, more elegant codebases, and accomplishing outcomes that reasoning and spray and pray methodology wouldn't enable.

The next step is to broaden my input sources. I'm thinking to create a new sceptic for each source, as quality will vary and different types of scepticism will be necessary to maintain operational rigour. I'll keep a human in the loop for the foreseeable future - sorting through all the brilliant ideas inspires my imagination and gives me ideas on new things to build to test capabilities.

I'm thinking to open source the tool but I want to hear from you first - If you think this could be useful, message me on LinkedIn or shoot me an email at contact@henryguyver.com.