I maintain a directory of AI tools in my knowledge base. Last count: 139 entries, each one researched, categorized, and dutifully filed. You would think a person with a 139-tool directory runs a spectacular AI stack.
The stack I actually use every day is six tools deep.
That gap is the most honest thing I can tell you about AI in 2026. Almost everything demos well. Very little survives contact with real work. So instead of another ranked list of tools I tried once, here is what actually runs my working day – and what each piece had to do to earn its place.
The Vault: Obsidian
Everything I do starts as a Markdown file in an Obsidian vault. Blog posts, product listings for my hat store, client strategy docs, a joke database, the knowledge base itself – all of it plain text, all of it linked, all of it mine.
This is the unglamorous foundation the whole stack depends on, and it’s not even an AI tool. That’s the point. AI models come and go; my notes have outlived several generations of them already. Plain text files on my own disk are the one layer of the stack that can never be discontinued, repriced, or acquired.
The vault matters for a second reason: it’s what my AI actually knows. Every tool below is only as smart as the context you can feed it, and a decade of structured, interconnected notes turns out to be a much better diet than whatever a chatbot can scrape in thirty seconds.
The Agent: Claude Code
If the vault is the foundation, Claude Code is the workhorse. It’s a terminal-based AI agent, which sounds intimidating and is actually the least intimidating way I’ve ever worked with AI – you tell it what you want done, and it goes and does it, showing its work as it goes.
What “done” looks like in practice, from my own recent history: auditing 175 blog posts for broken links and fixing them through the WordPress REST API. Filling in missing SEO metadata across an entire back catalog. Finding every reference to a defunct company across my site and cleanly removing them. Building product listings. Running SEO audits. Each of those used to be an afternoon; now it’s a conversation.
Full transparency: a lot of the routine writing on this site moves through this stack too. The direction, the opinions, and the final edit are mine – but I’d be a strange advocate for AI-driven content systems if I didn’t use one.
The Engine: SIE
The Strategic Intelligence Engine is the tool I couldn’t buy, so I built it. It does two jobs. First, it syncs the vault to WordPress – I write Markdown, and pages, posts, and product content flow to the right sites with the right fields. Second, it gives each site a chat assistant that answers from my actual content, using retrieval-augmented generation: Pinecone for the vector search, Claude for the reasoning.
It runs across my whole portfolio – this site, an ecommerce hat store, even a pizza shop. Same engine, different knowledge. That’s the quiet lesson in it: in a world where everyone rents the same models, the defensible part is your own data and the plumbing that puts it to work.
The Publishing Layer: WordPress
Twenty years in marketing technology and my first language is still WordPress. It’s fashionable to treat it as legacy, but it remains where the actual publishing happens – and its REST API is precisely what makes the rest of this stack possible. When my agent fixes a link or my engine ships a product description, WordPress is the thing gracefully accepting the request.
A stack full of frontier AI resting on a platform from 2003 sounds like a punchline. It’s actually the design principle: put the new, volatile intelligence on top of the old, boring infrastructure that has survived everything.
The Instruments: Google Analytics and Search Console
Not AI, still essential. The data decides what the AI works on. A recent 90-day GA report showed my humor archive quietly out-drawing everything else on this blog – so the humor library became a priority, and several ambitious plans for other content quietly did not. Without the instruments, an AI stack just helps you do the wrong things faster.
The Bench: ChatGPT and Gemini
Every craftsman keeps a second opinion around. Different models have different blind spots, and when something matters, I cross-check. That’s the whole role: useful, occasional, not load-bearing.
What Didn’t Make the AI Stack
The other 130-odd tools in my directory, mostly. Not because they’re bad – many are excellent – but because my rule for the stack is simple: a tool gets in when it removes a step from a workflow I already run. Not when it demos well, not when it trends, not when it promises a workflow I might someday want.
That rule is why the stack is small, Markdown-centered, and frankly a little boring. Boring is what reliability looks like from the inside. The fireworks are in the directory; the work gets done in the stack.
Adam Bernard is a digital marketing strategist and SEO specialist building AI-powered business intelligence systems. He’s the creator of the Strategic Intelligence Engine (SIE), a multi-agent framework that transforms business knowledge into autonomous, AI-driven competitive advantages.
For a stretch, NotebookLM was the tool I opened most. It answers only from the documents you give it and shows you the passage. That one rule is where it clicked for me that context, not the model, is the product. Here's what it does, who it's for, and what I did with the lesson.
-
Every AI chat that answers from your own content has a vector database somewhere underneath. I run one behind five sites. Here is what it does, why it is the last thing my engine checks, and the maintenance problem the tutorials leave out.
-
An agent is only as good as the tools it can call. I build MCP servers and run a WordPress adapter that hands Claude about thirty of them. These are the six rules I follow now, each one learned from a tool that went wrong.
-
In my AI stack post, Google Gemini got half a sentence. It deserves more. Here's the honest account of a model I respect, use occasionally, and haven't made load-bearing - and why it might be the right starter for you.
