I've been working with AI tools for a while now, and I wanted to share some thoughts on what's working, what isn't, and where I think we're headed. This isn't some grand theory about AGI or the future of humanity - just my opinions and experiences from someone who uses these tools every day and has opinions about what makes them actually useful versus what makes them annoying as hell.

FWIW - My perspective is definitely that of a power-user. I use AI to write code and research pretty much for like... 8-10 hours a day, every day. I recently started working on an AI-SaaS as a solo dev and so I have been very, very busy doing a ton of tasks with AI agents. So I understand that my opinions may not be shared by everyone, but hopefully my perspective can provide some insight to a broader audience nonetheless.

Tools That Really Get It

I have noticed a pattern in that the AI tools I actually enjoy using have something in common. They're very clearly designed by people who understand that AI isn't magic. They know it has limitations, they know it can fail, and they design around that reality instead of trying to hide it behind fancy marketing speak and "revolutionary" features that don't actually work. They put the power into the user's hands, and they make it easy to control the AI to get the best results you can.

There's definitely a skill to using AI effectively, and I think this is where a lot of people get tripped up. It's not about starting your prompt with "you are an expert..." (seeing these prompts make me cringe tbh). It involves understanding what these tools are good at, what they struggle with, and how to structure your work and data to play to their strengths. Like, Claude is fantastic at reasoning through complex problems but can get weird about certain types of creative writing for example.

I think that the tools that empower people who've taken the time to learn this stuff are the ones that will actually drive value. I believe that the tools that try to make AI "accessible" by hiding all the complexity will end up being worse for everyone - beginners don't learn how to use AI properly, and power users get frustrated by the lack of control.

I'm a big fan of explainable, simple AI experiences. Show me what's happening under the hood. Let me see and control the full context window. Don't try to be too clever behind the scenes. Just give me the tools I need and get out of my way.

The Editor Wars

Code editors are ground zero for AI integration right now, and it's where all the key patterns are being established. This is probably the most interesting battleground in AI tooling.

I started with VSCode and Copilot, which was fine but felt pretty limited at the time (they've since improved it quite a bit). Then I switched to Cline for their superior agentic mode, then to Roo a little while after. I preferred these experiences over Copilot for a variety of reasons, but two big ones being the ability to stop generation on a dime and ability to checkpoint and backtrack in a conversation. At the time, these were two huge advantages for Cline and Roo. They all worked well enough and integrated with my existing workflow and extensions.

I switched to Zed about a month and a half ago, and this is where things got interesting for me. I LOVE the performance of this thing - being built in Rust really shows, especially when you're working with large codebases or monorepos. But more importantly, the agentic editing feels like a genuine step forward. All these little features put you in complete control of the AI interaction.

You control your context window directly. There's no magic "infinite context" with hidden vector indexing like Cursor and Windsurf do. Instead, you explicitly choose what files and context to include. You can start new threads from summaries of old ones, which is brilliant for when a conversation gets too long but you want to maintain the important notes for a new thread. There's an "Open Thread as Markdown" feature that shows you exactly what's being sent to the model - user messages, agent responses, attached file contexts, everything. You can tag context and force tool calls with "/" commands. Being able to focus on errors and warnings with /diagnostics is also excellent.

A look at the Zed agent mode.

It's these little features that give you both transparency and control. I feel like it's a true AI connoisseur's tool - something that takes a little bit more effort to use, but lets you do exactly what you need to do. It doesn't try to be magic; it just gives you really good tools to work with AI effectively.

I actually tried Cursor and Windsurf when Zed had an ESLint issue for a week and I needed something that worked immediately. I only played with them for a weekend, so take this with a grain of salt, but they were both really disappointing imo. In Cursor, I couldn't even figure out which model I was using, which I didn't like. In Windsurf, I was doing some agentic coding and the STOP button became unclickable while it was changing a bunch of files. I literally had no recourse besides force quitting the app. I've never uninstalled something so fast in my life. If you're going to have the responsibility of having near full control over my code-base, you better let me stop you from doing something dumb.

I am very curious about asynchronous remote agents, but also I am a bit skeptical about their reliability. I think they have a giant UX problem in the fact that their value prop means that your entire workflow becomes just reading and reviewing diffs all day. I don't think any developer is really interested in that sort of workflow. I tried ChatGPT remote agent once for a few tasks and while it did complete them, I just wasn't like... totally sure how their changes would fit in with the rest of my code? I struggled with even coming up with good feedback because it was just so much content. I ended up just leaving the PRs open for a few weeks and then eventually closing them.

Interesting chat between Steve Yegge of SourceGraph and Nathan Sobo of Zed where they talk a bit about the UX challenges around diffs.

DeepWiki

DeepWiki has become a core part of my daily workflow, and it represents a perfect example of what I mean about tools built for people who understand AI's strengths and limitations.

I use DeepWiki to find targeted solutions to tough coding problems. It has full codebase indexing and search for public repositories, which is incredibly powerful when you're trying to understand how some library actually works or find examples of specific patterns. The key insight is that I can ask it questions about codebases without polluting my main context window where I'm actually working on my problems, and this is super wonderful.

Like, I might be working on a React problem in Claude, but I need to understand how some specific hook works in a popular library. Instead of cluttering up my main conversation with research questions, I can go to DeepWiki, ask it to explain the implementation, get the answer I need, and then bring that knowledge back to my main problem. It keeps everything clean and focused.

I experimented with running DeepWiki as an MCP server so I could access it directly from Claude, but frankly that defeated the whole purpose. Those conversations would pollute my actual problem-solving context, and suddenly my coding conversation would be full of random tangents about library internals.

So now my main workflow is to first work through the research problem with DeepWiki, then paste the solution into my main coding agent with a note like "Another agent provided this as context." It's a bit manual, but it keeps everything clean and purposeful. I'd rather have two focused conversations than one messy one.

MCP

Speaking of MCP, I honestly haven't found it solving real problems for me yet. I've experimented in Claude with MCP servers for Filesystem and Github access, but frankly I get more frustrated than anything when my Claude starts making random tool calls. I'll be trying to solve a specific problem, and suddenly Claude is off reading random files or making API calls that aren't relevant to what I asked for.

I want to provide context myself to solve problems, not have my model start making arbitrary calls to random services. I like to know exactly what information the AI is working with. When it starts pulling in random stuff from my filesystem or making Github API calls, I lose track of what it knows and what it doesn't know.

Maybe this reflects just my perspective. I use AI as a pair programmer and task completer, which shapes what I want from my primary AI service. I want it to be really good at thinking through problems with the context I give it, not trying to be autonomous and figure everything out on its own.

However, where MCP does shine is in secondary AI tools. I set up a monitoring agent for our Polaris web application, and this thing is fantastic. It tracks server usage and VM performance, I can chat with it via Discord to get quick updates that would normally take me longer to investigate myself. I have monitoring CRONs that alert me when thresholds are exceeded. It's like a super easy to configure observability dashboard that I can actually talk to.

This is the PERFECT use case for MCP imo. The AI agent has a specific job (monitoring my infrastructure), it has the tools it needs to do that job (server access, metrics APIs), and I interact with it when I need information. It's not trying to help me code or think through problems - it's just really good at one specific task.

I just don't see myself using MCP for my main coding or chatting workflows. I'd rather piece together context myself through copy-paste and file sharing than have all this tool-connected complexity. For primary AI tools, I want simplicity and control. For secondary tools with specific jobs, MCP makes total sense.

I know there are some folks out there that are basically nesting MCP servers inside of MCP servers for wild fully automated, multi-supervisor, microservice-like workflows. I haven't seen any of these setups yet, but I know that this is a thing that people are experimenting with. Curious, but I have my reservations about whether these setups are scalable or reliable.

A video outlining the microservice MCP servers idea.

Mini-rant about ChatGPT...

I just have to say that I absolutely hate using gpt-4 in my coding tools.

It's comical how it gets completely stuck on some of the most simple problems that Claude can handle without breaking a sweat. Like, I'll ask GPT to help me debug a simple React component and it will do some of the most inane shit. Function calls called wrong, "// .. the rest of your code" deleting 95% of my files, . Meanwhile, Claude will immediately be like "oh, you're missing a semicolon on line 23" and actually solve the problem I asked about.

I'm curious about GPT-5, but everything I'm hearing makes me think it's going in the wrong direction. Nate B Jones, a YouTuber whom I respect (he has great takes on AI stuff), keeps talking about how GPT-5 requires a lot out of you but it's worth it. He's been sharing these goofy prompt hacks and tricks to get the LLM router to send you to the right subsystem. Stuff like "Ask it to think harder" or "Tell it to use its advanced reasoning mode."

I just find this stuff to be absolutely ridiculous and it's a terrible sign of things to come. If my AI tool (which is supposed to make me more productive and smarter) is asking me to learn a bunch of arcane prompt engineering spells just to get it to behave properly, then frankly it's not worth the trouble. AI is already arcane enough - adding additional layers of complexity is not going to make anything better. Good tools should make things simpler, not more complex. I don't want to have to become a prompt wizard just to get my coding assistant to write decent code.

The whole thing feels like we're moving backwards. Early GPT was simple - you asked it something, it gave you an answer. Now we're building these complex systems with routers and subsystems and special commands... I want my tools to be smart enough that I don't have to be too smart about using them.

Building AI is Mostly Just... Building Software

As far as my experiences go in building real AI agents, it's still mostly just data science and standard web development. You're cleaning data, indexing it, optimizing retrieval, handling authentication and security. The AI part is actually the easy bit - it's all the boring infrastructure around it that takes time.

The agents themselves are pretty easy to build and extend. I honestly worry that Silicon Valley is overfitting to the complexity here. You'll see these conference talks about "advanced agentic architectures" and "multi-modal reasoning pipelines," but when you actually build this stuff, it's usually just some API calls to an LLM with a decent prompt and some basic error handling.

The actual development process for these agents is generally straightforward. Sure, there is still a lot of work in optimization and fine-tuning (getting the prompts right, handling edge cases, making sure the agent doesn't go off the rails), but getting a first-pass agent working is quite easy. You can have something functional in a weekend if you're not overthinking it. I am personally working with LangGraph a lot, but I am completely in the phase where I can recognize that the framework is so heavily opinionated and I am feeling that pain quite a bit.

The real challenge involves creating, curating, and managing the massive amounts of data these agents need. This is where most of the actual work happens. You need to figure out how to ingest data from various sources, clean it, chunk it appropriately, embed it, store it efficiently, and then retrieve the right pieces at the right time. It's a lot of normal data engineering.

On this front, I'm really excited about what Turbopuffer is doing. They have some unique approaches to vector storage that I think will stand the test of time. We're using it for Polaris and loving both the speed and cost. The pricing is also incredibly reasonable - like, suspiciously reasonable.

They have some unique and interesting approaches to leveraging cloud object storage as the means to store and retrieve data. This allows for absurd scale that databases like Weaviate or Pinecone simply can't match just due to core infrastructure limitations.

A wonderful podcast featuring the CEO of Turbopuffer, Simon Eskildsen, where he discusses the challenges in scaling vector storage.

Creative Applications Are Thriving

Where AI really shines is in creative applications with proper constraints. I'm particularly interested in projects like Weavy, which gives you fine-grained control over AI image generation. It has the power of something like ComfyUI but with way more polish and usability. You can actually understand what's happening at each step instead of just throwing prompts at a black box and hoping for the best.

A look at Weavy's composition node which allows you to create complex AI videos from multiple input sources and gives you full creative freedom over the output.

The key insight here is that creative professionals don't want AI to replace their creativity - they want tools that amplify their vision and make it easier to execute on their ideas. Weavy gets this. Instead of trying to be "AI that makes art for you," it's "AI that helps you make the art you want to make."

EVE Online YouTubers are doing some fascinating work using AI to bring their stories and characters to life in ways that weren't possible before. They're using AI-generated voices for different characters, creating custom artwork for their narratives, even generating background music that fits the mood of their stories. But the key thing is that they're still doing the creative work - the storytelling, the character development, the narrative structure. AI is just helping them execute on their vision more effectively.

YouTuber Warlock Industries uses AI voiceovers to bring life to EVE Online characters.

When creators use AI as a tool to amplify their vision rather than replace their creativity, the results can be genuinely impressive. But when they let AI do the thinking for them, it shows immediately. You can tell when someone has just fed their script to ChatGPT versus when they've used AI to help them polish their own ideas.

Where We're Headed

I think the future belongs to AI tools that respect both the user's intelligence and the technology's limitations. AI is normal technology, and the best tools will be transparent about what they're doing, give users control over the important decisions, and focus on amplifying human capabilities rather than trying to replace them entirely.

The tools that try to hide complexity behind "magic" interfaces are going to lose out to the ones that empower users who take the time to understand how to work effectively with AI. I think we're going to see a split in the market between "AI for everyone" tools that are simple but limited, and "AI for power users" tools that require more knowledge but give you much more control and better results. I am personally much more excited about the latter.