The context is everything for AI and Humans

Just a few months ago I wrote about the AI bubble. That was a take from a financial perspective, I still stand by it, but only idiots never change their minds. Since then, my regular use of Claude, Codex, OpenCode and Pi has only increased I can now count a dozen use cases that I could never live without again.

Making sense of LLMs, or the importance of context

It all clicked for me when I discovered a paradigm that has changed my approach with work. Context is everything in human relations, and so it is in neural networks. At the beginning of the "AI era", most users including me, started with Hey Gemini, what's the weather like on Sunday? How should I dress? type of questions.

That didn't really scratch the surface, and returned a general sense of meh.

It all made sense when I started to work for the AI agents providing them with enough context to become useful. Looking at the Claude desktop app, it's clear how the Anthropic team designed the client to steer the user into working in a specific way: Cowork is for working on files in a specific local folder, Code is for coding via terminal and git, chat is for everything else.

What's impressive to me is that by adopting best practices in real life, artificial life gets better the same way. This is a paradigm shift. Computers used to be programmed by adopting a language, methodology and organization that is the computer's, not the human's. With AI agents the well-organized individual can succeed at tasks that were unimaginable, while learning how to be organized, how to tidy up ideas, how to document thought, how to summarize concepts into actionable specifications.

AI use cases that can inspire

My work with AI agents started with folders of files providing context to chats. Things started getting better and I could see improvements each day I tried more complex operations.

I immediately realized that I needed a stable organization method, a persistent data storage, reusable primitives and mental model. I could benefit and gain from it.

I set out to build my knowledge base spanning across work and non-work items, and the work is in progress. I didn't decide to turn myself into a robot, I just decided that I could benefit myself by doing the creative work of thinking, writing, documenting, summarizing and then organizing the information I need. Because then, a computer would act on my behalf quickly and precisely.

Today I am connected to several Directus instances via the MCP server. Both Codex and Claude normally retrieve updated data and are able to provide me with insights, reporting on the meaning of changes in data that a human doing queries on a database would not immediately capture. This is a feature that itself can augment my capabilities, giving me time to do the creative and speculative work that humans excel at.

I have coded several new web applications that I wanted to do for years. Some of them are live, some are work in progress, regardless I have been increasing the rate of learning through complex problem that I could solve thanks to LLMs general knowledge and prowess at coding, not to mention in managing infrastructure.

My Wrangler and Github CLIs in my terminal are there sitting idle and authenticated, whilst my Claude checks my unused Cloudflare workers and spins up a Hyperdrive to serve data to an Astro.js website deployed on Cloudflare pages, all done by Claude.

The pre-requisite for succeeding with AI

I think it was The Primageon on YouTube who once said that AI is a multiplier. A bad engineer, or net-positive -2 engineer, will produce -20 work if given AI. A good engineer, or a 10x engineer, could get 100x the output. I believe it to be on point. The human is still using his own knowledge, experience, reasoning and creativity while operating AI agents. I could analyze, maintain and configure AWS, Cloudflare, Vercel, VPS servers, Postgres or Clickhouse databases before AI; Today, I can do it 10x faster and 2x better as I know what's happening, I just need it done, double-checked, and monitored in time by an agent that I have instructed to be able to move to something else.

What I'm saying is that there is no escape from skills and experience. I'm 47, I have had my experience building career and I then started applying automations and AI to my problem solving. For those starting, there is no shortcut to learning, doing things manually, reading the notes and understanding the documentation. If you are given a calculator and have no math knowledge, I doubt you could go past additions while making any sense.

Vibe coding

The term vibe coding clearly refers to producing slop code that is poor quality, risky, ill conceived and potentially dangerous. It's often associated with coding the entire code using AI agents rather than using agents for certain areas while maintaining full control of the codebase. I think this definition should be revised.

There is no harm in coding entirely via AI agents. The dangers are in the additional powers that non-technical, inexperienced users can today deploy to production changes that no experienced engineer has conceived or reviewed. The slop is in the output phase, not the generation. I consider myself more of a Technical Product Manager than an engineer; my understanding of code and infrastructure is advanced, but I don't code for a living. I am entitled to make decisions as Technical Team Leader, yet I don't code more than 20% of my time. This gives me the right to deploy, because I know the process, the risks, the mitigations, the concepts of web apps, languages and infrastructure. If I decide to not write any code and let Codex do it, I know what I'm doing.

Vibing an app isn't a sin. It's an opportunity. You can learn a lot by doing it, you will also learn from your pain when you underestimated security checks and got hacked. It's all part of a process that's just faster than before, but to me looks exactly the same.