Metadata →
Claude Code has six levels that show how to use it better, from simple prompting to managing multiple sessions. Beginners start by giving commands, but progress by planning, managing context, creating workflows, and scaling projects. Mastering these levels helps users avoid mistakes and improve their work with Claude Code. [!note]
Este es un excelente recurso para estructurar un curso de uso de IA, dirigido a personas que están en el nivel 1 para llevarlas al nivel 3.
Highlights
id996042024
And then here, down at the bottom of the status page, you’ll see all of those agents. So, at main, is like the actual terminal you’re talking to. And then we have the three sub agents, and they’ll be color-coded. So, I can hit the down arrow, and I can kind of switch between them.
id996035468
level one of Claude Code is the prompt engineer. Now kind of funny when you think about it, cuz a few years ago, this was like a legitimate job title. Like that was the upper echelon of AI practitioner. Was a prompt engineer. And now, a couple of years later, it is the lowest of table stakes in terms of how we interact with agentic coders. Now, what does it mean to be stuck at this level, and why should you care? Well, you should care because the outputs are the very definition of mediocre, right? This is where we get generic AI slop. And the reason you’re getting generic AI slop is because your relationship at this stage is very one. Way. You treat Claude code as a blunt instrument, as a simple tool. It is not a collaboration, and it is not a partner. And because of that, you hop into the terminal and you just start giving it instructions, things like build me my AI agency website
Prompt Engineering es la forma más básica de interactuar con el modelo, implica tratarla como un instrumento desechable, en donde le doy una instrucción y espero que mágicamente desarrolle lo que necesito, sin tomar en consideración todo el contexto del conocimiento implícito y los supuestos que no he compartido con el modelo.
id996036275
tell Claude code to be adversarial, right? You can even get kind of complicated with this. And when you have a build code, say, “Hey, I want you to spawn essentially like a, you know, devil’s advocate sub agent that’s taking a look at everything we’re building and telling us what’s wrong with it or if there’s a better path.”
Esta estrategia, propia del nivel dos en el framework del autor, marca una transición desde el uso de una herramienta al uso de una relación, en donde yo le pido al modelo que asuma ciertas funciones para identificar aquello que no soy capaz de ver, para que piense junto a mi.
id996038469
If you were starting a project or even like a brownfield project, if you did /init inside of Cloud code, it will essentially take a look at how you’ve written all the code, create the Cloud MD file and say like, hey, here Are the conventions based on what I’ve seen? However, we are now seeing studies done like this one last month where things like claw.md and agents.md are actually harmful to these coding agents. In fact, across multiple coding agents, they found that context files like that tend to reduce task success rates compared to providing no repository context while also increasing inference cost by over 20%. So what does that mean for us? It means the skill of knowing what context cla can benefit from and what will only hurt it is one we need to master. And oftentimes the answer is cla code needs some context, not too much context. And oftentimes less is more.
Interesante, lo que se declara aquí es básicamente aplicar la noción de “content rot” a CLAUDE.md, lo cual implica que “menos es más”, y que una habilidad a desarrollar es entender qué tipo de contexto es necesario y cuál no.
id996039558
loud code is so strong. If you are just some monkey on the keyboard who just hits next, next, enter, enter. Yes, yes, yes. You will still be able to build a lot of stuff. You truly will because Opus is that good. But the problem occurs when we eventually need to do something that’s more complex. And this goes back to my idea of like you don’t know what you don’t know. If you start dealing with complex projects, you are going to need to start understanding at a fundamental level what the building blocks are. What is a front end? What is a backend? What is authentication? What are databases? How does security work? How do I deploy this? You don’t need to know how to write any of those things. But you do need to know how all the building blocks come together. And again, you don’t need a CS degree to do. That. You literally just need to ask cloud code why and how. Right? If cloud code does something that you don’t understand, whether it was a fix for something or a decision, you need to ask these questions. And one of the, frankly, the best use case for AI is the infinitely patient tutor, because I can ask it to explain something to me in 3,000 different ways, 3,000 different times until I understand that. But to do that does require a mentality shift. And that’s something we’ve talked about throughout all the levels, right? You need to take an active approach and an active role in how you talk to Claude code
Interesante consejo. Claude cada vez toma más decisiones de manera inteligente y uno queda sin entender las decisiones que toma, lo cual es caldo de cultivo para la generación de deuda técnica y cognitiva. La idea de preguntarle es súper buena porque te permite aprender en la práctica misma en base a un tutor infinitamente paciente.