How to be a good AI driven coder?

By TheBlindGuy07, 20 November, 2026

Forum
Assistive Technology

So first I am putting this in AT cause the app dev and prog forum asumes apple related, mine is general.
TLDR. I am a bad programmer, I know I'll improve over time. But I am an even worse vibe coder / ai assisted programming person, and I just don't know where to start. The tech always changes 6 months ago and even poor fireship has difficulty keeping up with the new trend lol. Joke aside, I am well aware about the risks and the financial instability of all this mess.
But... since 2022, my first reflex like everybody was to copy paste into chatgpt. I knew github copilot existed after chatgpt and even then I hardly tried github copilot inside vscode or visual studio. Then came anthropic with npc that killed most of the weird over the weekend pluggins, which was imo a good thing as it became the standard, as far as I know. Just like openai endpoints are, whatever it means. I know now that I have enough junior knowledge, and especially, enough exposure from my immersion in the dev / tech culture for more than a decade at this point, to be able to search the right thing pre llm era, and now, to look for the right first path to try post llm. I know that 99% llm is maybe lucky with the first happy path, and as the project grows, if you don't have the full mental picture you will be brainroting your way in the llm until skill atrophy puts you so down that you'll be worse than before you used any llm. And I have some ambitions, so I know I want to avoid this path. Generally speaking anything beyond the web / mobile / desktop crud, and anything science or math that's, well, brand new, is almost impossible to get right with llm if you don't pour a significant part of your own efforts and energy into it at the very least.
Considering this, how can I best use LLMs? Like... At my scale of ideation projects, even 1m context window is overkill. I have started to use claude desktop to experiment and only in my happy hellscape do I truly understand the expression "vibe coding". :) And I use claude desktop on windows cause even blind people like pretty navigeable UIs.
But I know this is only the beginning. I understand agents and orkestrator, but don't know how to use them. The add context button in all ides... Like do I add it myself or llm can infer automatically? How is it different from memory. I understand skills, connectors and etc. But as the project grows and become worthy... do each folder / feature set has its own md file telling the llm what it is so it doesn't dump the full codebase in it? Do we do this ourselves? Or it's both us and llm that can help us speedrun that. ... How and where to start? I really hope my question is clear cause I don't even know how to formulate.
Not to mention all the llm technical jargon itself and how to implement it in apps or use the raw api as any sain skilled developer does to work with llms.
PS: no llm has been used at any point in the above text.

Options

Comments

By Choom on Friday, November 20, 2026 - 06:00

I've been thinking about giving AI assisted coding a try, with a local model of course since I don't like to depend on third-parties for coding, and am wondering about the same. That said I still hope that this bubble bursts; NVIDIA, which was the first company to reach the $5 trillion market cap, is now down to $3.52 trillion at the time of this comment. On the other hand, Anthropic's IPO just valued a company that has only losses to report at $2 trillion, so the bubble may take a while to burst and I want to work again.

As I understand it, AI assisted coding is all about providing instructions in plain English; even skills are all about that, like scripting, in plain English. Then there's the Model Context Protocol, which provides functions with descriptions also in plain English that explain to models how to interact with external services thus making them agentic. I just don't know what software to use, on macOS specifically, to take advantage of all this. I believe that OMLX and LM Studio both provide MCP and skill support, but LM Studio is not very accessible in my experience, and I'm yet to try OMLX. Finally there's editor integration, which requires an MCP-aware editor that acts like an agent; Xcode can do that I think, and so can Cursor, but the latter is forked from Visual Studio Code which just doesn't work for me on macOS so Cursor might suffer from the same problems.

I can't help here as I know nothing about this but wanted to put my thoughts out there anyway.

I've just started using Claude and like it but that's just for simple projects.

I don't think these things are going anywhere, the names and tech might change but the idea of typing into an agent in simple language? That wil just keep getting better and better. I'm happy about that honestly as it will make our lives easier in the long run.

By Blind AI on Friday, November 20, 2026 - 12:49

I've been in computing for about 40 years and was a systems analyst, back in the days of SSADM and Jackson Structured Programming. My honest advice: don't try to be a better junior programmer. Become the systems analyst and let the agent be the programmer.
That's the job that matters now. The agent can write the code. What it can't do is decide what should be built, how the pieces fit together, and how you'll know it works. That's analysis, and it's where your skill should go. Do that and you won't lose your skills. They move up a level.
In practice:
1. Write down what you want before any code exists. What the program does, who uses it, what goes in and what comes out. Plain English, a page or two. That's your spec.
2. Keep a briefing file in the project. Claude Code reads CLAUDE.md and Codex reads AGENTS.md. It says what the project is, how it's organised and any rules. You start it, and the agent helps keep it up to date. Ask it to update the file whenever something important changes. When the project grows, add smaller ones in subfolders. So, the answer to your question is both of you.
3. Make the agent explain itself. After each change, ask it what it did and why, in plain words. If you can't follow the explanation, stop and ask until you can. That's how you keep the whole picture in your head.
4. Get tests written early. Ask the agent to write tests that prove each feature works. Tests are how a blind analyst checks the work without reading every line.
5. Use an agent, not a chat window. Pasting into ChatGPT is like phoning a programmer and reading them your code. Claude Code or Codex works inside your project folder, reads the files, runs them and fixes them itself, and both can work fine from a terminal with a screen reader.
Don't worry about the jargon. MCP, connectors and skills are just plug-ins and instruction sheets for the agent. You'll pick them up as you need them.

A machine was involved. It is much smarter than it was forty years ago!

One major problem for me is that, unless I'm running the models locally, I am growing a dependency on a service that's almost impossible to replicate and is offered by just a few companies at a loss.

I also don't regard system analysis as an extension of development but rather as a dimension orthogonal to development, because in my experience, as a developer I should trust the analyst's skills and not need the full picture in order to implement an API, and as an analyst I should trust the developers with figuring out implementation details. If developers start peeking into architectural decisions, or analysts start peeking into implementation details, they are micro-managing, which is bad.

By Blind AI on Friday, November 20, 2026 - 13:30

Fair points, Choom, and I think we agree more than it looks.

On dependency: that's exactly why I'd put the effort into analysis. A spec, a briefing file and a set of tests belong to you, not to OpenAI or Anthropic. They work with any model, including one running locally. If the big services vanished tomorrow, the analyst's work would move to whatever's next. The code-pasting habit wouldn't.

On orthogonal: you're right that "up a level" makes it sound like a ladder, and it isn't one. But in a human team those are two people. With an agent, one person holds both roles, and your rule is the one I'm recommending. The analyst trusts the developer, here the agent, with the implementation. You don't read every line. You check it against the spec through tests.

Asking the agent to explain what it did isn't micromanaging. It's the analyst confirming that what got built is what was specified. That's the analyst's job, not the developer's.

Same machine involved. Still smarter than forty years ago.

By Travis Roth on Friday, November 20, 2026 - 13:57

I agree with the above premise about junior programmer. The irony in this is the AI using companies want senior engineers still, and it is a bit of a mystery how the next generation of senior engineers will be developed. Anyhow that is above my pay grade..
It has been recommended to me by some pretty good senior level dev that the Claude Academy is a good place to learn. I haven't done the lessons as I feel my current knowledge is kind of in the middle somewhere and I like books more than online courses but perhaps when I get some motivation. If you are starting from scratch like the original poster it would be worth a look. (Note I'd venture a guess OpenAI has something similar but I've not searched.)
https://academy.claude.com/

By Choom on Friday, November 20, 2026 - 16:00

I think that the best way to understand how all the pieces fit together is to implement my own harness with Model Context Protocol, skills, and connectors support to run with local models like GPT-OSS-120B and Qwen3.8-Flash-Next. The harness can also offer some agentic stuff like the abilities to edit files in a local repository and talk to git out of the box. I'm not sure what I'll be getting myself into, but think that it will be a fun way to at least distract myself for a while.


Editing to fix the spelling of Qwen.