For the few past years, coding has changed. People went from typing everything themselves, to using Stack Overflow, then to ChatGPT copy-and-paste, then to autocomplete in the editor, and now to agents. We have come so far.

As we reach this point, we may have lost a few people, but this is your guide to coding with AI. This will explain the current landscape, the terms you need to know, the tools, and last but not least, what I think is next.

Type of coding

In recent years, there have been different types of coding, but right now I want to talk about 3 types:

Vibe coding
Augmented coding
Agentic coding

There could be some left out, but these are the ones you need to know.

What is vibe coding

Vibe coding is coding without looking at the code. Most vibecoders use tools like Chef by Convex, Lovable, Bolt.new, and many others. These are people who do not read the code and may not understand the code.

This is the most common type for non-coders. It is a nice way to get into coding, but I do not recommend you stay here.

Why?

Because this may have security vulnerabilities that will be hard to fix in the future, and you may not be the best person to prompt the AI for better results. The models are good, but if you cannot drive them in the right direction, you will not have the best result.

Imagine the model is just a smart child. It lacks common sense. It is lazy. It likes using common things. You are the parent and must show it where to go.

But I just want to add: if you know how to code and you are just vibe coding because you do not actually care much about the project, because it is one of those things you do not plan to maintain, then go ahead. You have the knowledge to drive it well.

What is augmented coding

Augmented coding is coding, but AI helps.

This is what early GitHub Copilot was. It was just about people coding with tab complete and maybe copying and pasting from a chatbox. It could be ChatGPT or my current favorite chat app, T3 Chat.

You are typing the code, but the AI helps. You may feel a small productivity boost, but not much, because you are still typing the code, so you are still the bottleneck.

So who should exist in this realm of coding?

I think people who are learning new technology. Because it does not stop you from using AI, but it stops the AI from replacing your learning and still lets you move faster.

What is agentic coding

This is a form of coding where you are not the one writing the code, but you review and test the code.

The people here usually understand coding and have experience, but they realize that them typing the code is the bottleneck for everyone. So instead, they prompt the AI on what to do.

And yeah, this is where I think the vibe coder should upgrade to. If they have understanding, can review and steer the model to the right path, then they can also review the code, see mistakes the AI makes, and tell the AI better alternatives instead of letting it take the easiest path.

That can make the code better or cheaper to run, which only benefits you either way.

So this is the realm that most serious people are in. People like Theo Browne, Ben Davis, Simon Willison, and funny enough, DHH, the Rails guy, code with AI now.

So for those who say AI cannot do serious coding: the creator of one of the most influential tech stacks in the world, the T3 stack, uses AI; and the creator of one of the most important frameworks also now uses AI. Then there is Ben Davis, one of my favorite dev YouTubers, who also uses AI.

So if you think AI is not useful, just click off this post.

Some terms

Now we know the types of coding with AI. Here are some terms you need to know.

What is a harness

Now that we know the different ways of coding with AI, here is one term you need to know: harness.

A harness is anything that gives AI tools. It could be anything: OpenCode, Claude Code, Codex CLI, Cursor, and many more out there.

But what is a tool?

A tool is a function that AI can call, and when it calls a tool, it is called a tool call.

But if you know anything about AI, you know AI can only generate text — at least models like GPT-5.4, Opus 4.6, and Gemini 3.1 Pro, which are some of the best models out right now. But they can only generate text.

So how do they call tools?

Using a special syntax.

<|start|>user<|message|>What is the weather in SF?<|end|><|start|>assistant<|channel|>analysis<|message|>User asks: “What is the weather in SF?” We need to use lookup_weather tool.<|end|><|start|>assistant to=functions.lookup_weather<|channel|>commentary<|constrain|>json<|message|>{"location": "San Francisco"}<|end|><|start|>assistant<|channel|>analysis<|message|>The weather in San Francisco is sunny. I can answer.<|end|><|start|>assistant<|channel|>final<|message|>It's sunny in SF.<|end|>

This is an example of a tool call using OpenAI Harmony format. In this example, the model is calling a weather tool to find the weather in SF.

But tool calls and harnesses are not complex. They just need 4 basic tools to become a coding agent:

a read tool
an edit tool
a list tool
and the most important tool, a bash tool

Why is the bash tool so important?

Basically, it can replace all the other tools just by using bash commands. It can use cat to read files, sed to edit a file, ls to list files in a directory, and many more. It can run any command that your system has.

With the 4 tools I mentioned, your agent can run commands, read files, list files in a directory, and edit files.

If you want to see a basic harness, click here:https://github.com/Jemo69/harness

And if you want to learn how to build a harness, click here:https://www.mihaileric.com/The-Emperor-Has-No-Clothes/

This will let you build your own harness and show how AI harnesses are simple.

What is an Agent.md

Now that we talked about harnesses, we need to talk about context.

The AI uses its tools to find context, like when it reads a file or gets a list of files in a directory. So basically, context is information about your project.

There are different ways the AI gets context:

the model provider’s base prompt
your harness’s system prompt
the AGENTS.md or CLAUDE.md file
and then your prompt

But AGENTS.md is a file in the project that the AI sees before getting your prompt. So a good one can help a project, but a bad one can confuse the AI.

So here is a piece of advice that helps when working with AGENTS.md:put things in it when the AI makes a mistake or forgets something. It should not be a large file with information about your whole codebase, because the model has tools for that.

Imagine if you were coding and your boss kept telling you about tests for a feature you are not working on every 30 seconds. You would lose your train of thought. That is what a bad AGENTS.md does. It confuses the model.

And models like the Gemini models really like to use any context you give them, so you might steer the AI in the wrong direction with a large AGENTS.md.

What are skills

But there are other ways to give your AI model context without making a large AGENTS.md.

And what is that?

Try a skill.

This is a large markdown file that a harness loads when needed, but it is not always in context. If it is not needed, it never loads. But once it is needed, then it gets loaded.

An example of a popular skill is Anthropic’s frontend-design skill. This is a skill that lets the model create less generic UI. It gives design capabilities to models that do not naturally have it.

Without the skill, Gemini creates better UI than Opus. I hope we can all agree that is true. Gemini may suck at tool calling, but at UI, Gemini is the best without skills.

But with skills, Claude is better than Gemini.

So you see, skills help.

And no, skills are not exclusive. A skill is just a markdown file, so anyone can use it with any model.

Another popular skill is GStack. I cannot give a review about that, but I have an opinion. Feel free to message me on X @JemoLife0213.

What are MCP

Now we have talked about all the ways to give context using markdown, but there is another way, and that is through the Model Context Protocol, also known as MCP.

While markdown is good for giving context, that data is mostly static. MCP is a way of giving AI models access to external data that is usually always updating. It is also good for adding extra functionality.

MCP has tools, resources, and prompts. These are the 3 primitives behind MCP.

Let me break it down:

a tool is a function that AI can execute
a resource is a read-only data source
prompts are instructions used to steer the model in the right direction when doing a task

Now you may be asking why people use skills when MCP exists.

Because an MCP can bloat the context with tools your AI does not need at that time, which can make results worse. That is a major disadvantage of MCP.

But like everything, a few may work. When it is too much, it becomes toxic.

What is a Ralph loop

Now we have left context and started to talk about how to handle long-running tasks.

First, I just want to say: with the current generation of models, Ralph loops are not needed. But nobody knows where this industry is going, and Ralph loops may come back or stay dead. Still, it is good to know what they are.

Ralph loops are named after the character Ralph Wiggum, who is simple-minded, naive, and persistent in The Simpsons. And guess what? These are the principles of Ralph loops:

simple: one long task, repeated in a loop
naive: the AI can make the same mistake over and over again
persistent: the AI will not stop until the task has been done

I think that is all you need to know about Ralph loops.

Best practices

Some good practices:

Keep context small. Do not feed the AI the entire codebase. Instead, just give the key details it really needs.
In AGENTS.md, do not put file paths everywhere. Just put things the model finds confusing or keeps making mistakes on.
Always have examples ready to be shown. In a SvelteKit project, I have demo.remote.ts files. This teaches the model how to write remote functions, which is a new primitive and only recently got added in the model’s data. But with good examples, the model will produce good code.
Keep the codebase clean. Do not pile in tech debt, because the model could use it as an example, and it will be a bad example.

What do I recommend

For a harness, I use OpenCode, but you can also use Codex, Claude Code, or put them in one app and use T3 Code.

And if you say you do not want to leave the terminal but still want the advantages of T3 Code, you can use T1 Code, which was made by a member of Theo’s team called Mariah.

For models:

GPT-5.4 and Opus 4.6
if you are willing to suffer a bit, try Gemini 3.1 Pro
for fast models, I recommend Gemini 3 Flash and GPT-5.4 mini

These fast models are for quick fixes, not large refactors or complex features.

For autocomplete, use whatever comes with your IDE. And if you use Neovim, I recommend Copilot and Supermaven. Supermaven is no longer in active development, but it is still a good option if you need tab complete.

If you want to change IDEs, I recommend Zed and Cursor.

That is it for AI IDEs.

There are other things too, like the Codex app and T3 Code like I mentioned.

What is next

The way AI is going, nobody can predict the next 6 months.

I just want to say I think multi-agent setups will come up more. It could be through worktrees, or it could be through cloud agents with their own computers.

And I think more review and testing will be automated to give the AI more of a feedback loop and make the AI’s results better.

Conclusion

So now you are ready to start coding with AI.

And remember:

augmented coding if you are learning
agentic coding if you want to ship fast
vibe coding if you do not care

And remember, you can reach me on X @JemoLife0213.

And remember: stay coding.