Oct '25 • Newsletter Discussion
My Complete AI Stack Evolution - Part 1: The Core Stack
Over the past year, I've tested dozens of AI tools. Most of them are gone now. Not because they were bad—many were technically impressive. They're gone because they didn't survive the test that actually matters: did I still use them six months later when the novelty wore off and real work pressure returned?
This is a three-part series covering what stuck, what didn't, and why the difference matters for anyone building their own AI workflow.
---
The Core Shift: ChatGPT → Claude (But Not Really)
---
Most people frame AI tool selection as an either/or decision. ChatGPT or Claude. Gemini or GPT-4. As if you need to pick a side and commit. That's not how it actually works in practice.
I started with ChatGPT like everyone else. It was and remains an excellent tool. Custom GPTs work well for specific workflows. The prompting framework is solid. The integration with other OpenAI services is seamless. I still use it regularly, particularly for quick queries and tasks where I don't need deep context.
But this year, I found myself defaulting to Claude more often, initially for writing. Claude's tone control let me sound like myself rather than producing the generic ChatGPT cadence that everyone's learned to recognize. Then they introduced MCP integration, and everything changed.
MCP—Model Context Protocol—isn't just another feature. It's a fundamental shift in how AI assistants interact with your actual work environment. Instead of operating in an isolated chat window where you copy content in and copy results out, Claude with MCP connects directly to my file system, my databases, my actual workspace.
---
What MCP enables in practice:
Direct file access means I can ask Claude to review a document, and it reads the actual file rather than requiring me to paste the contents into a chat window. That sounds like a small convenience until you're working with large files, multiple related documents, or iterative editing where context matters.
System integration means Claude can read from and write to Obsidian (where I keep all my project notes and writing), Notion (where I manage client work and publishing), and even my email. The AI works where I work, not in a separate application that requires constant context switching.
Workflow embedding means the tool fits into my existing processes rather than requiring me to build new processes around the tool. I don't change how I work to accommodate Claude. Claude adapts to how I already work.
The technical capabilities of ChatGPT and Claude are comparable. Both are excellent language models. But the integration layer is what makes Claude indispensable for my daily workflow. When I open my laptop in the morning, the computer actually feels like it's there to help me accomplish things rather than being another interface I need to manage.
This doesn't eliminate the need for proper setup, though. Claude has specific instructions configured in my app preferences like so:
---
- all of my project knowledge and writing is in Obsidian, which you have access to via MCP. Always check in there
- ask clarifying questions before giving answers
- do not make assumptions. If you've been given a link, you must access it, or tell me straightaway if you cannot.
- do not conflate or exaggerate any facts
- do not tell me everything I do is brilliant, or be a sycophant. I expect logical, factual challenges.
- My LinkedIn Profile is https://www.linkedin.com/in/jim-christian-digital/
- My homepage is: https://jimchristian.net
- My Newsletter homepage is https://signalovernoise.at
---
And I have TextExpander snippets configured that I use periodically throughout the day to log my activity in Obsidian. That way I can review what I've done and ask Claude to reference specific work later when context matters. Everything still needs proper direction. The AI doesn't magically know what you want—you have to build the systems that make it useful.
The larger pattern here: model capability will continue to rise, fall, and plateau across all vendors with increasing frequency. That's expected in a rapidly developing space. The services that win long-term will be the ones that integrate most closely with what users actually need rather than the ones with the most impressive benchmark scores.
---
Search Intelligence: Perplexity + Pages + Comet
---
Perplexity replaced Google for research about a year ago. The shift wasn't immediate—it took a few weeks of using both in parallel before I realized I'd stopped reaching for Google automatically. What made the difference wasn't just that Perplexity provided better answers (though it did). It was that the entire interaction model fit how I actually think about research.
With Google, I had to structure queries. Think about keywords. Consider how other people might have phrased the question I was asking. Evaluate which search result might actually contain the information I needed versus which ones just had good SEO. That's cognitive overhead that adds up over dozens of searches daily.
With Perplexity, I ask questions the way they occur to me, and I get answers with actual citations to real sources. Not links to pages that might contain relevant information somewhere. Actual answers with the specific passages that support them highlighted and referenced.
I still remember the moment that made me realize this was different. I was trying to fix a problem with my car door—something about the locking mechanism not engaging properly. I asked Perplexity a conversational question about it, and the results included a YouTube video showing exactly the repair process for my specific car model. Not buried on page two of search results. Not something I had to click through three different forum posts to find. Just directly relevant, immediately useful information.
Perplexity Pages is an often-overlooked feature that lets me create knowledge bases on specific topics. When I'm researching something complex that requires synthesizing information from multiple sources, I can build a Page that collects everything in one place. It's perfect for creating quick reference guides or sharing research with clients without having to format everything manually.
The Comet browser has become my default on Mac. It's an agentic browser—meaning it actually does things for you rather than just displaying web pages. When I need to reference something I was reading earlier, Comet can find it. When I need to extract information from multiple pages, Comet can handle that. It sits in the emerging category alongside Dia and the Claude plugin for Chrome, and it's showing the most promise in what's still a very new space.
I fall back to Arc on Windows and iOS because Comet doesn't have versions for those platforms yet. But the pattern is clear: standard browsers plus standard search equals constant context switching. Comet eliminates that friction by bringing search intelligence directly into the browsing experience rather than treating them as separate activities.
---
Why Integration Matters More Than Capability
---
Looking at Claude and Perplexity together reveals the pattern that determines what sticks: integration with existing workflow beats raw capability every time.
Both tools could be technically superior or inferior to their competitors on any given benchmark. That matters less than whether they reduce friction in actual work. Claude integrates with where I write and think. Perplexity integrates with how I research and browse. Neither requires me to change my fundamental work patterns—they enhance the patterns I already have.
That's the difference between tools that get abandoned when the novelty wears off and tools that become indispensable over time.
Questions to ask about your core tools:
- Does this tool work where you already work, or does it require switching contexts?
- Are you copying data between applications, or do they connect naturally?
- Six months from now, will this still reduce friction or just be another thing to manage?
1
0 comments
My Complete AI Stack Evolution - Part 1: The Core Stack
powered by
skool.com/signal-over-noise-1446
Practical AI implementation space. Real examples, direct feedback and no hype. For people building systems, not theorising about them.
Suggested communities
Powered by