November 20, 2026

What AI Does in My Content Process, and What It Doesn’t Do

Black and white comic ink drawing of a writer handwriting a note at a desk, with an AI agent on the computer screen.

I make blockchain content for a living, and I use about 7 AI tools to do it. I’ve written 2,000+ articles and have hosted 170+ podcasts, 110+ X Spaces, and 80+ on-the-ground interviews at crypto conferences and events.

I use AI tools in different ways and for different capacities. They range from self-hosted agents to agents on frontier models to familiar tools such as ChatGPT and Claude. Each one handles a different job, which I’ve slowly added over the past 2.5 years since first using ChatGPT to help draft and edit articles.

The short version is this: “Today, I don’t use AI to create my content for me.” I give AI specific goals (i.e., outlines, objectives for copy) and deliverables I want (i.e., articles, video edits) that are hand curated at the very beginning of the process.

I then use AI to produce more content and edit more content. I don’t hand off any of the thinking or creative processes to AI. That is relegated to me, the human, for everything that people see.

Where it started

The first AI tool I used for content was a GPT built by my writing coach. He’d trained it on thousands of articles from our website, Neo News Today. That backlog of thousands of articles taught it how we format our pieces and how we sound.

Honestly, I resisted at first.

I had spent years learning to write, and I did not want to outsource my thinking. Then I tried it with my own work. I put in my own outline, my own information, and my own rough draft. The result looked very close to what my team publishes.

That was the first time I leaned into AI for my process.

Research, copy editing, and images

Today I use ChatGPT and Claude for general research and for line-editing. I send them first drafts, and they check grammar, spelling, and whether any thoughts are incomplete.

I also use ChatGPT for image generation. Many of the headers in my long-form X articles and newsletters come from prompts I send to it.

Private questions, and images other tools refuse

Some questions I would rather not send to a large model. For those, I use Venice. It is an aggregator of language models, built by the team behind ShapeShift, and founded by Erik Voorhees.

Venice lets you choose the model you want. It also keeps your information away from the model provider. Good examples of privacy-oriented questions might include those related to health or personal finance.

Or, if I’m being extra paranoid, about projects that haven’t launched yet, as Anthropic has settled for similar instances in the past.

Venice also generates images that ChatGPT or Claude will not produce, such as those that raise copyright issues.

Self-hosted agents, for builders

OpenClaw is a self-hosted agent service that runs on a separate Mac Mini. This method is a bit more complex, as you build your agents and their harnesses yourself. It is definitely the techno geek’s choice.

I use OpenClaw for personal searches and for things I don’t want to share with frontier labs models, like those developed by Anthropic and OpenAI. I also use OpenClaw for work purposes, where I don’t want my information used as training data. For example, when I’m writing articles about teams or projects building in the Neo ecosystem.

That’s the closest thing to proprietary data for our team, and the only place I want the internet or agents to extract that information is from our website, our go-to resource.

If you have never built a website theme from scratch, or never worked with code, OpenClaw may not suit you. My OpenClaw agents use Venice on the back end.

The weekly newsletter agent

Every Sunday, an agent scrapes an X list of more than 130 accounts. It then drafts my weekly newsletter. It is the first agent I use every week (or the last, depending on how you view Sundays). The list this agent uses is also the same list I use to monitor the blockchain ecosystem I work in.

Before I agentified my life, I was spending four to five hours every week putting this newsletter together. Now I spend about one hour a week putting it together. Most of that hour goes to review and formatting. The draft takes a few minutes.

This is by far the biggest payoff working with AI has given me, in terms of getting my time back.

Two agents for articles

The article pipeline uses two agents, one for research and one for drafting. It was important to create two distinct agents for this process because I wanted each agent to do a phenomenal job at its assigned task, not to have one agent be okay at two different assignments.

Generally speaking, the more specific a task that an agent is assigned, the better it will be at performing that task to your standards.

The first agent, Ghanima, is named after one of the twins from the Dune series. It does research. It works through GitHub repositories and websites, then turns the findings into a Markdown file that another agent can read.

The second agent, Leto, writes the first draft. Leto’s voice, tone, and style were trained on 20 of my best articles, dating back to 2018.

I am proud of this setup because I spent a lot of time refining Ghanima and training Leto. In terms of deliverables, it is the equivalent of having an intern I onboarded and taught to write.

The output is a draft I can review, edit, and ultimately publish on our website.

Transcription on the desktop

Transcription is the second thing I am most proud of. Frankly, one of the killer uses of AI will be the forthcoming wave of people who type less and speak more. By speaking to an agent or a computer, a human can think, speak, and direct more naturally. This is going to be a huge unlock.

I speak into a microphone and record a file on my desktop, drop it into a folder, and it gets transcribed. That transcription is then sent to an Obsidian Vault and organized according to the project or topic I was speaking about.

I use this method for podcast interviews, Spaces recordings, and team meeting minutes. I also use it when I onboard a new client, to build a profile and a list of next steps.

A personal wiki

Each transcript feeds into a personal wiki, which serves as a knowledge graph. It holds the ideas, concepts, and themes I have been writing about for more than five years.

The result is a record of how my views on blockchain projects, the industry, DeFi, and so much more have changed since 2018.

Tools for people who will not build

Not everyone wants to build a self-hosted system, and that is fine. I loved building mine, because I want to own my data. That being said, it took many months of work before my workflows were reliable. I also felt like I lost a little bit of my mind while working 12+ hours a day to get them up and running.

For folks who aren’t interested in that, there’s GrokBot, Muse, and Dots. Those are “personal AI agents” that do just about everything OpenClaw agents do, except they’re not run on a local computer.

I use GrokBot mainly for social media, to grow my X account, and to study other X accounts. It feeds directly into X and uses the Grok model.

GrokBot also helps me follow changes to the X algorithm, research topics, and draft short-form video scripts.

Codex and Claude Code

For coding, I primarily use Codex because I think it is fantastic for this purpose. It helped with the heavy lifting on my recently re-released website. It also has helped me build small side projects, like a highlighter app for live streams.

I use Claude Code heavily for much of my video and audio content because I prefer stronger models for these purposes. For instance, Claude Code runs the Waterfall Content process, which helps me record and edit long-form videos and then extracts other forms of content. With the waterfall process, I record a 15- to 30-minute video, upload the raw video, and a set of agents extracts posts, short clips, and copy from the transcript.

I generally run any code that Claude Code generates past Codex afterward, as a way to steelman what I’ve built. This is especially important for non-developers who begin vibecoding things.

I’ve come far, but have a long way to go

I want to be honest: this article and the previous one, How Content Marketing Works in a Blockchain Ecosystem, are only the articles I’ve written using the transcript and copy from a long-form video. The processes are rough. My agents and I still have a lot to learn about setting up processes and making them more efficient.

That’s why I am building in public, so readers and listeners can see the wins and the losses as they happen. Hopefully, this is a source of inspiration for others who are interested in using AI tools in innovative, unique, and fun ways.

I got AI-pilled in February of 2026, when I hosted David Johnston from Morpheus on the Smart Economy Podcast. After that conversation, he got me jazzed about building a self-hosted agent. Before that, I used only the GPT my editor built.

So after eight months, this is where I’ve landed.

Coming full circle, I don’t use AI to create my content from scratch. I use it to produce more content, edit more content, and create more creatively. My AI agents only produce great content when they’re given a specific outline and goal to produce a deliverable that’s good enough for human eyes to read.

← Back to writing