DEV Community: L Black

The Internet of Intelligent Things

L Black — Mon, 09 Jun 2025 19:45:04 +0000

What happens when agentic AIs visit websites more than humans do?

According to Malwarebytes, 50% of all internet traffic is automated. As AI adoption grows and more and more time-consuming digital tasks are offloaded to AI agents, that 50% will inevitably increase. That thought should spark fear (or excitement) in the hearts of those whose business model revolves around display ads and click funnels.

A new paradigm is emerging that is going to bury those who don't adapt. At least, that's my take. It's difficult for me to see it any other way, unless AI adoption suddenly goes into decline, which seems unlikely.

People Doing Less

I told a friend once that AIs will write code and software engineers will just orchestrate them like a conductor. This was after seeing GPT-2 for the first time, when it was still relatively new. In about half a decade since then we've gone from inconsistent and incoherent chatbots, to AI systems that can build entire applications, do deep research across the web to produce detailed and coherent reports, as well as offer non-technical support and advice on all manner of topics and situations.

The growing capabilities of AI have obviously not gone unnoticed. Consumer adoption has been unprecedented. ChatGPT launched in late November 2022. By the end of January 2023 it already had 50M users. It's currently at around 800M. Enterprises were slow off the mark but are rapidly trying to become "AI-first" as they battle to get ahead of—or, at the very least, keep up with—competitors.

First-Hand Experience

I see this growth when I look at the adoption in my own social and professional circles too. And my personal experience, of course, is a prime example. Since I joined ChatGPT back in 2022, I've interacted with AI on a daily basis. Now, I'm using multiple AI services daily. From coding and writing tests, to research, design and ideation, I let AI do the heavy lifting whenever possible, so that I can focus on the more fun and interesting stuff.

I'm seeing the same behaviour in others, and it's no wonder, because it makes sense. Why would you spend hours or days traversing the internet for information when AI can do it for you in a few minutes? Or trawl through documentation for some very specific piece of info that AI could probably find in seconds?

Another example, just to drive the point home: I came up with a novel design for a piece of jewellery and I wanted to find out how much it would cost to get it made. I put not one, but two AIs to work to go do the research, considering different metals and techniques to achieve the structure and mechanism, distance from my location, quoted lead times and, of course, prices.

I used Manus and ChatGPT's Deep Research. Both spent about 10-15 minutes visiting numerous sites, extracting and evaluating data, compiling it, and eventually provided me with all the details I needed. If I'd done the research myself, between work, projects and other commitments, it would have been a minimum of days, but most likely, weeks.

Inadequacy of the Website

While 15 minutes to do weeks' worth of research is a massive productivity boost, it could still be better. Our websites were built (often poorly) for humans to navigate, not machines. Sure, we can add schema markup and other metadata to make it more machine-friendly, but the AIs still have to:

wait for a UI to render
hunt through navigation menus
deal with pages heavily dependent on visuals
find data scattered across multiple pages
deal with cookie banners, modals, and other intrusive UI elements

Dealing with only one of the above is the best-case scenario, but access could also be entirely blocked by CAPTCHA-like tests of humanity. The bottom line? Our websites are not optimised for AI.

I think there's some irony here that the best way to improve the human experience is to improve the AI experience. And this is why I think a fundamental shift is coming to the web. Two emergent characteristics will change the nature of websites to such an extent that I think we'll probably need to use a different term for these "neosites". The emerging characteristics? AI-Accessible, and AI-Enabled.

A Brief Evolution of the Web

We started with static websites in Web 1.0.
We made them dynamic and data-driven in Web 2.0.
Smartphones came along and our attention turned to mobile-friendly and mobile-first.
Alongside the focus on mobile, we saw the emergence of the Internet of Things, real-world devices connected to the internet.
Finally, we had the rise (and stall) of Web 3.0, which sought to bring decentralisation, blockchain, and data ownership to the forefront.

Now Artificial Intelligence is poised to unlock characteristics that will transform the internet again, but more dramatically than ever.

AI-Accessible Websites

For proof of what the future holds, we can look to the partnership between OpenAI and Shopify. ChatGPT, which has had product recommendations for a short while now, will gain access to Shopify's new Catalog, giving it real-time product information from millions of Shopify merchants. This type of integration means you could discover and purchase products online without ever leaving chat, or even opening a browser.

While OpenAIs Operator could theoretically hunt for and purchase products on your behalf, it's an inferior experience, because Operator:

requires a $200 per month subscription
uses human-style browsing to navigate the web and take action
still requires manual input for payment details, etc.

The UI-based data discovery is the big bottleneck here, which the Shopify integration obliterates. Because Shopify has made its data AI-Accessible, ChatGPT can query it directly, bypassing search engines, UIs and clunky navigation for quick data retrieval and display. Waiting time is slashed. The human experience immediately gets an upgrade.

This is AI-Accessible in a nutshell: the site provides AIs direct access to rich, structured data; the AIs decide how best to display it to their users.

AI-Enabled Websites

At Microsoft Build 2025, Microsoft announced NLWeb, an open source project that:

[...] aims to make it simple to create a rich, natural language interface for websites using the model of their choice and their own data.

I tried it, and integration was...not simple. Still, while NLWeb might not yet work as advertised, it's an indicator of their vision of the future. I think they see what I see, and that's intelligence baked directly into websites. NLWeb seems to be a tool on the road to making that a reality.

But what does AI-Enabled give us? It gives us websites that don't just provide data, they contribute to problem-solving. In a sense, ChatGPT was the first true AIE website. Our current and most common experience of AI-Enabled websites is general purpose AI chat services, like ChatGPT, Claude, Gemini, DeepSeek... Of course, AI is the core value proposition for these services.

We also have the more traditional "dumb" chatbots, which were previously more annoying than useful. A site with a dumb chatbot doesn't qualify as AI-Enabled. But we're at a point in the evolution of AI that intelligent chatbots could become integral to elevating passive websites to something far more valuable.

Let me give a simple example. After receiving the reports from Manus and ChatGPT, I was left with a couple of short lists of jewellers, including links to their sites. It would be amazing to be able to visit one of the sites and ask questions specific to the jeweller, including current lead times.

Adding this type of intelligence to the site would be beneficial to the jeweller too. The AI could handle all the repetitive queries, potentially qualify leads, and even begin the conversation and extract information about design preferences and materials.

Or consider a non-technical user who needs to buy a new laptop. Instead of scouring the internet trying to figure out what to buy, they could go to the online retail site of their choice and simply start with, "I need a laptop". The AI would take the conversation from there, asking them about the purpose, budget, operating system preference, then present them with the best options. Low stress, high efficiency.

Neosites

The point here is that the web will shift away from being navigation-driven (search -> click -> nav -> click, click, click). The web will be intent-driven, where the human expresses their intent and AIs act to satisfy the intent. The companies who succeed in the "neoweb" (continuing the "neo" theme) will build intelligent sites that can respond to intent.

This becomes even more powerful when sites are both AIA and AIE. Theoretically, sites could make the "intelligence" accessible through an AIA interface or endpoint. This would allow AIs to communicate directly and work with each other to solve the user's problem.

At Google IO 2025, Google announced the A2A (Agent-to-Agent) protocol. It's designed to allow:

[...] gen AI agents, built on diverse frameworks by different companies running on separate servers, to communicate and collaborate effectively...

The infrastructure to allow "intelligent things" to communicate and collaborate across the internet is already being built. This is not theoretical or speculative. It's already happening. If I wasn't sure before, the recent moves by Microsoft, Google and other players give me more confidence than ever.

A new web is coming, and it will be defined by these two characteristics:

AI-Accessible: "Here's my product data, you figure out what to do with it"

AI-Enabled: "Tell me what you're trying to accomplish, and I'll help you solve it"

The Fallout

All of the above leads to the overarching question: If an "internet of intelligent things" is the future, what impact will this have on the web as we know it? I have some ideas, but none of them will be surprising given the progression I've already laid out. For this reason, I'll keep it brief.

What Changes

Display advertising becomes obsolete, in most contexts. In a world where AI agents browse the web more than humans do, the ROI on display advertising plummets.

Click funnels collapse, for the same reason as display ads. The intent-driven web brings the data to the human. What might be producing sizeable revenue now, will generate only a trickle.

Search engines start to favour AI-accessible and AI-enabled websites. Human UX still has value, but AI usability is prioritised. We may also see search engines themselves become AIA. We've already seen the early version of AI-Enabled, with Google's AI overviews and "AI Mode".

Website traffic becomes overwhelmingly dominated by AI agents.

Brand expression, perhaps surprisingly, becomes more important than ever, because it has to permeate not only visual design but also structured data passed to AIs. Brands will need to figure out how to convey their essence in the age of AI agents and intelligent things.

Who Wins

Early adopters and first-movers. Those who build for the future, rather than focusing on what's "hot" today. Platform providers (like Shopify) who make their services AI-Accessible will see significant increases in revenue through partnerships and AI-driven discovery on the new web. In general, companies that evolve their products/services into intelligent service nodes.

Who Loses

Businesses built around or dependent on display ads or traditional funnels (obviously).

"Dumb" sites. With AI agents using AIA search engines that favour AIA and AIE websites, the traditional website will be buried in results pages. The intelligent neosites will reign supreme. A million websites that fail to adapt will fade to irrelevance.

What Remains

Human-centric platforms. Forums, communities, social media, and other spaces where the value is in human connection and shared ideas and experiences. Sadly, this is likely where display advertising will also consolidate.

Gaming platforms fall under the same umbrella. But things get interesting when generative gaming becomes viable and mainstream.

Underlying infrastructure, for the most part. AI may help us improve it, but the current emergence won't fundamentally change it.

Human decision-making. The human-in-the-loop remains, but supercharged with AI power.

What Now?

It doesn't matter if I'm wrong about every single thing in this post. The one thing that is undeniable is that AI is a revolution that is changing how we do everything. So, even if things don't change the way I suspect they will, we can at least be certain of one thing: things will change.

The question is: Will you be ready? Personally, I'm already building for the future. Are you?

Cursor Crash Course for the Uninitiated

L Black — Sun, 01 Jun 2025 07:32:17 +0000

Cursor is powerful IDE that extends the normal capabilities of the AI models you likely already use. It's built on a fork of VS Code, so if you've ever used that then the interface should feel familiar. I've only been using it for a few weeks, but in that time I've come to establish a routine that is consistently giving me higher quality results than when I started.

What We'll Cover

Start with a Plan - Creating detailed planning documents
Agree on a Directory Structure - Avoiding structural chaos
Commit (Often) - Using Git for safety nets
Rules, Rules, Rules - Configuring AI guidelines
- User Rules
- Project Rules
Beyond Code - Research and ideation workflows
What Else? - Community discussion

Start with a Plan

It's sometimes quite amusing to see how easily AI can deviate from what you've discussed, especially in a lengthy conversation or long-running task that eats up context. And it only takes a tiny deviation to quickly end up in hole, since AI can make changes in the editor about a 100x faster than a human.

I avoid this by creating a detailed plan, usually through natural conversation with the AI. I tell it my initial ideas, share any thoughts on technologies and architecture, and have a back and forth as I might if collaborating with someone a project. The details get drawn out in the conversation. Once I feel like we've covered all the key areas, I have the AI turn the conversation into a detailed planning document. Depending on the complexity of the project, a single document might be enough; in other cases, we (the AI) will write more detailed documents for individual aspects that require breaking down further.

Once the plans are written, I review. If there are errors, I correct them or tell the AI. These plans then serve as a consistent reference point that we come back to again and again, checking things off as they get done, or updating if we alter any decisions along the way.

A written plan is one of the most reliable ways I've found to ensure the AI doesn't get lost in the development process.

Agree on a Directory Structure

This might seem like overkill, but you won't think so when you're 50+ actions into your vibe-coded project and you're suddenly like:

Wait, why is there another src directory containing only services?

Did I mention that AI can make a lot of changes very rapidly? This is even more likely if you live on the edge like I do. I'll often just set it to work and walk away to do something else. But if you do this without agreeing on a clear structure for your project, you may come back to all sorts of gnarly decisions made in your absence.

Don't wait for the WTF Moment™. Agree on the project structure, then have it create it or write it to a doc—perhaps one of the planning documents you created earlier. Personally, I do both.

Commit (Often)

Before you put the AI to work on your detailed plan, I strongly advise that you initialise a Git repository for the project. Cursor has its own feature that creates "checkpoints" before each set of changes. You can then revert the project to any of these checkpoints if you run into problems (and you will).

But I've found it's fairly easy to lose track of changes when relying solely on automatic checkpoints. It's safer and easier to just commit regularly, especially after a feature has been implemented, a bug fixed, or some other satisfactory outcome from the conversation. If you want to guarantee you can always get back to a working state, commit early and often.

Rules, Rules, Rules

An important feature of Cursor is its Rules. These are basically prewritten instructions that can be sent along with requests. You can have both global and project-specific rules, as well have rules added to every request or to specific types of requests via pattern matching. None are configured by default, so it's up to you to add the instructions that matter. I find they're good for a variety of project preferences:

coding guidelines
testing strategy
commit frequency & style
tooling

You can find the rules in Settings > Cursor Settings.

At the time of writing there are two main sections.

User Rules: [...] sent to the AI on all chats, composers and Command+K sessions. Basically, they're global, persist between sessions and windows, and get sent with every chat message.

Project Rules: As implied, they only apply to the current project. You set them by creating .mdc files in .cursor/rules, either manually or through the Project Rules interface. A good place to put rules about project specific coding conventions. I usually add these files to my repo to make life easier for anyone else who might work on them.

Cursor User Rules

But what exactly should you put in these rules? That depends what matters to you. Here's what I put in mine.

# Rules of Development (RoD)

These are general guidelines to follow for writing, modifying or refactoring code.

1.  **Modularity:** Keep functions and modules small and focused on a single responsibility.
2.  **Organisation:** Adhere to the established folder structure.
3.  **Readability & Maintainability:** Write clean and understandable code. Only add comments where its purpose is not clear from the code and context (e.g. complex logic, counter-intuitive implementation).
4.  **Modifications & Refactors:** If you are proposing changes, always check with the user first before implementing, unless already explicitly told to do so. ONLY make changes relevant to the specific task at hand. 
5.  **User Confirmation for Installations:** Do not install any software, packages, or dependencies using terminal commands without explicit confirmation from the user first.

You'll notice that I've given them a heading. This allows me to draw attention back to them simply by punctuating instructions with something like, "Remember the RoD".

You might also have noticed that there aren't many of them. Perhaps it's just me, but I find the AI can get "overwhelmed" or confused when there are too many rules. This varies from model to model, of course, but I've found that less is more and that it's better to target specific things with each rule. Each of the above was added to address a specific undesirable behaviour.

As an example, before I included these explicit instructions in the rules, I found the AI was prone to creating massive functions and extremely long files. Classic Junior Dev. But now... It still creates massive functions and extremely long files—but it does so far less often ;-)

Jokes aside, I get far better results with the rules than without them. This makes them essential. Try for yourself and see what works.

Cursor Project Rules

It's a chore, but you should absolutely add rules for each project you work on. Just like in the real world, the rules and tooling often change from project to project. The Project Rules allow you to highlight the project-specific details so that the AI doesn't have to guess. Below are some things I've found useful to add in the project rules files.

Coding Guidelines

Any specific formatting preferences, I'll add here; naming conventions, indentation, curly brace position, take your pick. Linters can fix some of this stuff, but I'd rather start with something close to my preferred style than have to do lots of clean up later on.

Git Branching Strategy

Can vary between organisations, teams and projects. Whether you prefer Git Flow or some other strategy, if you want the AI to safely make changes to your repo, you'd better let it know how you like to manage it.

Git Commit Style

I started a project recently without any project rules, and was so confused when the AI was trying to cram a 10-line commit message into what should have been the commit title. Plus, it wasn't adding a subject. Of course, I realised the project rules were missing, so it didn't know that I prefer Conventional Commits.

Testing Strategy

I always include information about testing because AIs will often write massive amounts of code without a thought for any tests. And when you ask them to add some, they'll respond like, "Good idea!"

Deployment Strategy

Yes, I live so wild I'm willing to let AI commit, push and deploy changes to live environments. But only when I know it has the playbook on how to do so.

Here's a snippet from one of my project's .mdc files:

...
**Testing Setup**:
- **Framework**: Vitest across all packages
- **Workers**: `@cloudflare/vitest-pool-workers` with Miniflare for Cloudflare APIs
- **Environment**: Node.js for shared-utils, Cloudflare Workers environment for workers

**Testing Strategy**:
- **Priorities**: Unit tests prioritised over integration, comprehensive service layer testing
- **Creation**: Create or update unit/integration tests for all new or modified features and fixes
- **Frequency**: Always run all tests after completing changes

**Package Management**: PNPM with workspace dependencies (`workspace:*`)

**Deployment Strategy**:
- **Flow**: `master` → feature branch → work → commit → push → merge to master → auto-deploy
- **Staging**: Deploy branch directly before master merge
- **Production**: Master branch triggers deployment
- **Platform**: Cloudflare Pages (web), Cloudflare Workers (services)
- **Environments**: local, staging, production with environment-specific configs
...

While there's more detail in this rules file, each section and point is still highly focused. With this kind of detail provided, I've found the AI models I work with in Cursor are less likely to get stuck, or use the wrong tools or strategies.

Project Structure

Remember that directory structure I mentioned earlier? I'll often include it in a rules file. I haven't quite settled on a specific approach for this, I'm still experimenting with providing just a high-level overview versus a detailed file-level map.

One method I've used is to have a script that creates a directory tree diagram in plain text. I then have a pre-commit hook run the script and output the diagram to a dedicated rules file, which also gets added to the commit. This works well for smaller projects, but, as one might imagine, it's not so well suited for large projects due to the size of the map that is generated.

A variation of this approach could work, though, if the traversal depth and directory file count is limited. This would create a smaller map with enough detail to give the AI useful clues as to where to find/put particular types of files.

It's worth noting that Cursor did release a beta feature back in April that provides project structure in the context, though I've not seen that it actually works as intended.

It's Not Overkill

This might seem like a lot of extra work and support in order to get the desired results when using Cursor with your chosen model(s), but in reality it's not that different from working with people. All of the above are things you'd need to share, in some form or another, when onboarding new team members onto a project. Why would an AI engineer need any less?

Beyond Code

What has really surprised me about Cursor is how useful it has been beyond unlocking rapid engineering. I've increasingly been using it as both an ideation and a research tool.

Research

The true benefit of using Cursor for research is context. Whether you're starting with research, or in the midst of a project and needing to break out into research tangent, I've found it's both extremely convenient and effective to do it within Cursor.

To give an example, while experimenting with an idea I realised we needed a better approach to creating realistic skin texture for some aquatic wildlife. Because it already had access to overview and planning documents within the project, as well as all of the existing experimental code, I was able to set the AI on a research task with a single-sentence request. It searched several websites and forums, compiled the report in Markdown with links back to relevant sources, and added it to the project ready for viewing.

As I noted in a previous post, I also use Manus and ChatGPT for research. But if I'd turned to them on this occasion, I would have had to answer follow-up questions and explain far more.

Ideation

When I'm starting with nothing more than a vague notion that I'd like to explore and flesh out, I create an empty directory and open it with Cursor. I start the conversation with that vague notion and let it unfold. What naturally emerges from this discussion are samples, prototypes, and—if the original nebulous notion starts to really take form—overview and planning documents, architecture diagrams, etc. The project grows organically, and whichever model I decide to use has full access to this history.

Knowing where we're going and where we've been makes the AI less likely to unintentionally revisit old and dismissed ideas. It helps the AI understand the why of the current project state. It's all about context.

What Else?

The routines I've described are just what I've learned so far. I'm always experimenting, tweaking, chopping and changing my approach. It's better not to remain too attached to one way of doing things, especially in the age of AI. And Cursor tend to ship new features at least once a month, so things you have to do manually today might be automated tomorrow. Ain't that the truth?

What routines have you developed when working with Cursor? I'd love to know what's working for you, so please share in the comments!

AI and the Copy-Paste Problem

L Black — Fri, 30 May 2025 18:05:10 +0000

As the number of my AI service subscriptions has increased, a workflow naturally emerged where I pit AIs against each other. This is typically when planning projects, but I also use the same process to evaluate and improve AI-written code.

It often goes something like this:

Gemini drafts a plan
Claude evaluates it
ChatGPT refines the final draft

Which AI I use for each task is irrelevant here, especially at the rate they release new versions, which frequently shifts their dominance in different areas. The important part is the unavoidable context switching, and particularly the text-jockeying that needs to take place to complete the flow. It was a minor inconvenience when I was only using ChatGPT and then discovered Claude. I'd compare outputs and see which I preferred for the given task.

But the number of AIs and tools I use on a daily basis continually grows. At the moment I use Claude, Gemini, ChatGPT, Manus and Cursor, sometimes for the same aspect of the same project. As a result, I can end up spending a lot of time moving outputs between tools. Also, too often I have to remind them of my preferences; be concise, offer useful critique, use 2 spaces for indentation, etc. This gets annoying.

This is, of course, because memory capabilities vary considerably between models. Claude and Gemini have none. Manus has a limited memory feature, akin to the memory snippets ChatGPT had prior to its recent upgrade, making it capable of accessing all past conversations. That upgrade makes a remarkable difference, creating a greater sense of continuity and an assistant that is actually getting to know you. But even here, the memory access is limited to specific models.

I'm looking forward to when it's a standard feature of AIs, as I believe it's the path to truly useful AI assistants. I was having a conversation with family a few months back. They're not in the tech industry and only really learned through me about a lot of the amazing things AI can do now. I enthusiastically shared the latest developments at the time, including how I'd been using AIs to help me build far more rapidly, and with far less actual coding. The conversation naturally shifted towards the future and I shared with them some ideas about how it—and our use—would evolve. For a long time I held the belief that it would become deeply integrated into our lives, yet seeing it manifest is still incredibly exciting and even surprising at times. I think inevitably it becomes not just deeply integrated, but deeply personal.

However, I don't think it will be a "one AI to rule them all" scenario. It's clear that I'm not alone in my use of multiple AIs, and I think this trend only continues as more models are released for specific verticals, with increasingly narrowing use-cases. I'm regularly using 5 AI services today. In two years it could be double or triple that. This exacerbates the Copy-Paste Problem even further.

Maybe I'm overthinking this, or overly sensitive to this perceived inefficiency. Perhaps others have found better approaches than what I've described above—I'd love to hear them. Either way, it's interesting watching the evolution of our workflows, and the exponential growth of the tools that are evolving them.