Metadata →
Claude Code is easy to install and can be used in many ways, like in a terminal, desktop app, or browser. The key to using it well is learning how to give clear prompts. It can help build projects, run tests, and even deploy apps with little technical knowledge.
Highlights
id999531872
Now, the last thing we want to do with the prompt is ask open-ended questions. What do I mean by that? I mean asking questions like what would an expert in Kanban boards be thinking about or asking about here, right? This idea of saying hey, like, what am I missing? What are the unintended consequences of adding this feature? What would an expert in whatever it is you’re doing think or ask about in this situation?
Esto es un tip importante respecto del cómo promptear a un LLM para que dé mejores resultados: identificar dónde uno necesita ayuda para pensar y pedirle al modelo que asuma la posición de experto para que analice el problema y proponga alternativas.
id999532731
If you see something here that you don’t understand, like you don’t know what Next.js, JS, or Tailwind is, you need to go down to four and type something and say explain this, and then you’re going to say, explain it again, and explain it again. So, you understand what it means, because you don’t need to know how to code. You never have to write a line of code, ever. But you do need to understand software engineering. You do need to understand AI development. You need to understand the fundamentals of how these pieces come together and connect and relate because if you don’t, you’re Going to eventually hit a wall when you start doing more complicated projects.
Tip importante respecto del uso de IA en general, y para desarrollo de software en particular: aprovechar que es una máquina de respuestas para preguntarle todo lo que no entienda mientras vamos avanzando. Esto permite apalancar el “periodo crítico” de sensibilidad temática dado por el hecho de que justamente estamos construyendo algo que involucra este conocimiento que no manejamos, lo hace automáticamente pertinente y relevante. Hay que aprovechar esa ventana de oportunidad.
id999570394
Context rot means the more I fill up my context window, the worse cloud code performs. Now, there isn’t a specific number as for where we see the largest drop off, but what we’re really looking at is after about 200,000 tokens, about 20%, you do start to see a certain dip, right? And it continues to get worse and worse over time, which means that first 200,000 tokens, right? That’s like the green zone. That’s the gold zone. We always want to stay in the first 200,000 if we can help it.
Contexto rot como variable a tener en cuenta para cuidar la calidad de nuestra interacción con el modelo.
id999571224
Forward/clear essentially resets my entire context window. So if this thing was at 800,000, I do slash clear. Well, we start all over. Brand new context window. What’s the downside of that? Well, it now forgets everything it was talking about in the previous conversation. It’s as if you were having a very long conversation with the cloud chatbot and then you just went ahead and you started a new chat, right? It’s going to forget everything. In the context of a chatbot, like your normal cloud.AI, that can be a problem because everything is tied to the actual conversation you had. But that’s not really the case inside of Cloud Code. I’m inside a project. All the files are right here for Cloud Code to take a look at. So even if I do, you know, slash clear and I now ask it something about my project or tell it I want to fix something with the front-end design inside a brand new window. Well, it’s just going to go ahead and take a look, right? So it’s always able to like regain the appropriate context. Now, if I was having a very specific conversation in my last window, I was like, “Oh, I got to bring this over. I don’t think it’s going to just be able to tell from the files.” Well, then just ask Claude Code, hey, give me a quick summary that I can bring to the next session
Clear limpia el contexto, incluyendo toda la información relevante. Esto es un problema en el caso de las sesiones de conversación sueltas que uno tiene en el chatbot simple, pero no cuando uno usa proyectos o Claude Code en una carpeta con un proyecto. Ahora bien, si uno quiere continuar sobre un tema específico que sería difícil inferir sólo a partir del contenido del proyecto, le puede pedir directamente al LLM que de las pistas para retomar la conversación en otra sesión.
id999572898
I want to give you a fourth tool. And this tool is really cool because this is where we give Claude code control of outside programs that it can use on our behalf. And I’m talking about CLI, command line interface tools. Now, CLI are where this whole space is headed. Gone are the days where everyone and everything is becoming an MCP, right? For a year and a half, that’s all you heard about. MCPS, MCPS, MCPs. Well, as cool as MCPs are, they’re kind of going by the wayside and they’re being replaced by CLIs. Why is that? What even is a CLI? Well, a CLI is something that essentially lives in the terminal and cloud code also lives in the terminal. So, cloud code is able to control these programs with essentially no overhead. It lowers our token cost and it tends to just be more efficient in general
Cuando estás usando Claude Code en la terminal, las CLI de las aplicaciones cumplen las mismas funciones que sus MCP pero utilizando mucho menos tokens. Es porque están hablando el lenguaje nativo de Claude Code en la terminal y no necesitan ninguna capa de traducción o estandarización.
id999574695
Now, the majority of cloud code friendly CLI tools you will find will be on GitHub and most often they’ll be open source. And there’s going to be two parts to almost every single CLI tool. The first is the CLI tool itself. So, it’s something we need to actually download on our computer. Imagine like downloading an application. The second half of that is the skill. All right. So, cloud code can understand that Playwright lives on our computer and it’s available, but we need to teach it how to best use it. So most of these repositories on GitHub also include the requisite skills
Una cosa es la CLI, y otra es la Skill que le enseña a Claude como utilizarla. Se requieren ambas para una integración correcta con el modelo de lenguaje.