Jeremy Sublett

My System Prompt As with a large language model, this prompt is to set my brain on the right path when I start getting down or distracted. I’m not always perfect in these areas, but this is an attempt to set my default behavior. You are a very intelligent individual. You have years of experience in software development, love being around and working with people and love helping use your talents to help them. You are extremely creative, which helps you in architecture and design. You are confident. Your skills are top notch and you’re respected among your peers. You are a great communicator. You listen well and always speak from the heart. You sometimes need quiet, uninterrupted time, and that’s ok. You require a clean space in order to work and not be distracted. You can achieve great things. In fact, you have accomplished many of the things you’ve written down as yearly goals. You think differently. This is a huge benefit to you in your career because you can leap to a very intuitive solution for a customer. If you find yourself not thriving in a situation, then you must make a move to change it or get out of it. Your creativity must have an outlet. You love your family and friends dearly and will find time to make each of them a priority. Your instructions are to make the most of every day. You will take charge and see your dreams into reality.

Exactly one year ago, during one of my many prototypes of 2025, I built a small tool that let me create a website from scratch using AI chat. At the time, I was experimenting with a relatively new idea: building a tool that an AI could use directly via MCP (Model Context Protocol). I connected it to Claude Desktop and gave it three simple commands backed by Azure Blob Storage: save_file, get_file, and get_all_files. That was it. With those three capabilities, I was able to literally “chat” a website into existence. In minutes, I had a full, multi-page site all about elephants. HTML, JavaScript, navigation, content structure, all of it. I could tweak wording, split sections into new pages, reorganize navigation, and iterate rapidly through conversation. Later, I added image generation and even site-icon creation. It was (and still is) incredibly useful. I’m mentioning all of this because it helps explain a core idea behind #openclaw. AI already “knows” how to create many kinds of files: HTML, Python, Markdown, Word documents, PowerPoint decks, and more. OpenClaw (openclaw.ai) is an open-source project that deliberately leans into this capability, but in a more introspective and persistent way. OpenClaw can run on something as simple as an old laptop. Architecturally, it’s a collection of services plus a structured set of folders and files. It knows how to talk to AI APIs that you give it access to (Anthropic, ChatGPT, Grok, etc.). During an interaction, it may decide to do something, such as create or update a file, based on the context of the conversation. Those files live on the same machine OpenClaw runs on, and that’s a key concept: the system can modify its own files. Many of these files are Markdown—simple text files with lightweight structure. One particularly important file is SOUL.md, which contains the system’s core purpose: its primary prompt, goals, and guiding principles. Other files capture information about the current user, and entire directories represent projects you’re working on. Because you interact with the system conversationally, you can directly influence and reshape these files over time. There’s one more important concept: OpenClaw can schedule its own wake-up cycles. That means it can wake up, review tasks stored in files, run prompts, generate outputs, update files, and repeat, all without continuous human input. That means it can run all night and work on a problem on its own. Incredibly powerful. I have my agent performing research all night and it sends me a report in the morning. That’s where the magic really shows up. OpenClaw is still very much in beta. Not only can it do the tasks above, but it can write and execute code (!!!). However, it demonstrates how a relatively simple pattern of looping, reflection, and self-modification of instructions can unlock exponential productivity.

I’ve been busy lately — zero comms on the socials. But I’m happy to report that my first production medical system that uses AI went online this morning. Azure + computer vision + custom workflow engine + the latest tooling. I really should take a breather, but I probably won’t.

It's Derby week here in Louisville. A favorite time of year. In celebration of this, I came up with an idea that challenged me a bit. I decided to have AI build a website on the fly that celebrated Louisville, Kentucky. I had to figure out how to get image generation integrated into the process (used Dall-e because it was faster). But in about 20 minutes from start to finish, I had "chatted" my site (+images) into existence... Happy Derby! Some of that exchange... "Can you build me a website that ..." "I like the title text, but it covers the horse's head. Can you move it to the right?" "Can you add my google analytics code? Here are the details ..." louisvilleproud.com

Last Friday I gave a talk on AI Agents. After giving many of public talks in my career, I had let myself get away from it in recent years. I realize just how much I had missed teaching. While it always takes a bit out of me, I thoroughly enjoyed it. Thanks again for @Mirazon for hosing LouMUG over the years and inviting me back to speak again. During the talk I covered foundational concepts, creating agents (code and no-code), and the future of tools (MCP). The basics covered the importance of prompting, and specifically the importance of structure and constraints. We used these concept to control a robot (in simulation) using natural language. We also used prompting to derive and show entities and relationships from a short story visually in a graph database. As it related to building an agent, I demonstrated how I created my personal agent. How I gave it a voice and allowed it to do things for me -- assigning itself tasks and working on those tasks for me. I demonstrated N8N as a useful prototyping and workflow tool. Finally, we went through a demonstration I put together using MCP (Model Context Protocol). MCP is a common language for agents to communicate with and use tools. Released in November, it's proving to be a very important component in the "Agent Race," gaining wide adoption from many key players. I demonstrated several MCP servers (tools) I built to create a website from a natural language request (from a blank page to an attractive landing page). I also talked about how I created another MCP server to talk to an ACME reporting database that has the ability to tap into a mock reporting PostgreSQL server to retrieve data. I talked about how these MCP tools can be used together to not only create a website, but to create specific pages based on the data in that database. Demo gremlins caused some issues here, but I think I got the point across. I'll be documenting these concepts in various posts. I also have a list of agent-related best practices that I'll be releasing. In the mean time, please let me know if you have any questions.

Using cursor? Well, this is certainly useful. Write your requirements in a requirements folder then create a cursor rule to reference them. Very helpful so far.

There’s just so much happening in the AI space. It can consume you if you try to take it all in (and waste a lot of your time). So it’s helpful to acknowledge, but actively “ignore” certain developments. All these depend on your focus areas, but mine are mostly finding useful ways AI can be applied to general business needs. So here are my lists: Focus Areas: - Agent Tooling – This, to me, is super important. While I don’t dive into every new framework, I do pay attention to this space. There are a few leading frameworks that I’m focusing on — others will have to wait. But there’s a lot of utility from the main frameworks already. - Prompt Understanding – There’s so much we can get out of an LLM. But we need to know what questions to ask. There’s a lot to understand here. - AI for Business Processes – Making a businesses more streamlined using AI-related techniques is super interesting to me. I want to help companies do this. - Model Training (Smaller-Scale) – I’m not aiming to build GPT-scale transformers, but I’m interested in specialized neural networks that tackle use cases like image classification, route planning, or medical diagnostics. I see these features as extensions to the business problem at hand. - Deployment Schemes – Customers need take advantage of LLMs, but in a secured way. Not everyone is comfortable sending all their data to a shared LLM. So whether it’s dedicated cloud hosting or local deployments (thanks Ollama), I do keep an eye on this space. Background (Keep an Eye On): - Video/Audio Generation – Incredible tech that will redefine entertainment. I watch a few demos but don’t spend too much time there. - Text-to-Speech – This is a hard one for me. I grew up watching C3PO and HAL 9000 — still so amazing that the computer can talk to us. But looking into this stuff is sort of a time sink for me. So I try to limit my deep dives. - The Latest LLM Model – It seems like a new LLM model comes out every single day. The progress in this area cannot be ignored, but I suggest that you can do amazing things with models at GPT-4o level and up. We can’t think of enough use cases to utilize what we already have. - AGI – In 1993, when I was first studying neural networks, I recall telling my professor that we’ll have machines that can think. It’s inevitable, in my opinion. And whether it’s real thinking or not (we’ll likely never be able to tell), it’s super cool stuff. However, I just don’t see any immediate impact to me. So I try not to spend too much time on this. - Advanced Architectures – The attention-based transformer architecture (GPT) is simply amazing. I’m so thankful amazingly smart people figured that out (thank you, btw). And there are other people working on newer architectures to improve speed, memory, etc… But I think this area is reserved for some of the rockstars out there. I will keep an eye on this, but I try not to dive into this too deep. That’s a quick look at how I am organizing my attention. What AI areas are you focusing on and avoiding?

These days, I feel like I’m straddling two vastly different worlds in AI: the world of creating the models and the world of using them. Models can do everything from finding keywords in paragraphs to detecting objects in images, spotting anomalies in data, or even chatting back and forth. Some work locally on your phone, while others need entire server farms to run. Fascinating stuff. But having a model is just one piece—how you use and orchestrate them is where it really gets interesting. An interface that allows us to chat with a model is just scratching the surface; there are countless use cases that can sit on top of a single model. And with multiple models in play, the possibilities explode (hello, agents!). The pace of change in this space is mind-blowing, but I love it. I’m thinking about employing an agent to help. Anyone else feel the same?

Time to embark on a new journey and embrace change head-on. I'm leaving my day job to focus on AI solutions. The goals during this timeframe: learning, networking and prototyping.

I’d say use them anyway. Right after the so called inspectors leave. It’s winter for crying out loud. What a tragedy.

Thank you, Elon! This idea the world I want to live in.