Stop Prompting AI. Use Karpathy's Method Instead.

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Stop Prompting AI. Use Karpathy’s Method Instead. · Channel: Larue · 06:08 · Watch on YouTube

Topics: AI agents, Prompt and context engineering, AI content generation, RAG, search and information retrieval

Summary

The video presents six ways to work more effectively with AI, attributed to Andrei Karpathy: provide rich context, create a detailed project spec, equip agents with the right tools, define success criteria, ask AI to challenge your ideas, and build a reusable knowledge base. It gives examples of applying these ideas to a salon booking form, voice-over generation, and research for future projects. The bonus advice is that while you can outsource thinking, you cannot outsource understanding: people still need to understand what they are building and evaluate the agent’s work.

Key points

Start with a spoken brain dump covering your goals, doubts, constraints, previous attempts, and examples; ask AI to reconstruct your intent and identify gaps.
Create a project spec that records the desired outcome and rules that must remain true.
Ask AI to interview you if you need help clarifying the goal, then save the goal and rules in a Markdown file in the project.
Give your agent the tools it needs to complete the task, such as a speech-generation tool for creating voice-overs.
State success criteria rather than prescribing every step, so the agent has room to find an effective approach.
Ask AI to make the strongest argument against your idea before settling on it.
Organize research in a knowledge base of Markdown notes that an agent can update and reuse.
Stay involved and understand the work: that understanding helps you guide the agent and judge its output.

Chapters

Start	Chapter	Summary
00:00	Brain dumps and project specs	The video recommends sharing rich, unorganized context and creating a detailed spec that records the project’s goal and rules.
02:33	Give AI tools and a finish line	The video demonstrates equipping an agent with a voice-generation tool, then advises defining success criteria instead of dictating every step.
04:05	Challenge ideas and build a knowledge base	Ask AI to argue against your ideas, and organize research in reusable Markdown notes that an agent can update.
05:37	The importance of understanding	The bonus advice is to stay involved and understand what you are building, because that helps you guide the agent and assess its work.

Quotes

The benefit comes from richer context, not messiness itself. — 00:30

you have to work with your agent to design a spec that is very detailed — 01:00

Don’t tell it what to do. Give it success criteria and watch it go. — 03:34

You can outsource your thinking, but you can’t outsource your understanding. — 05:37

Concepts and entities

Andrej Karpathy, brain dump, spec, WhisperFlow, Fish Audio, Codex, base de conhecimento

Full transcript (original language)

00:00 Andrei Karpathy is one of the most respected AI researchers in the world. He co-founded OpenAI, led AI at Tesla, and now works at Anthropic. In today’s video, I’ve packed six of Karpathy’s best tips that you can implement into your workflow to get much better results with AI. And at the end, I’ll show you the one bonus thing Karpathy says you should focus on in the age of AI. Tip number one is start with a brain dump. You don’t need to organize every thought in your head. You don’t need a perfect prompt. LLMs are incredibly good at

00:30 reconstructing long incoherent rambles. Karpathy’s advice is to turn on speech-to-text and just ramble for about 10 minutes about everything you’re thinking about, your goal, doubts, constraints, previous attempts, and examples. Then ask AI to reconstruct your intent and identify gaps. The benefit comes from richer context, not messiness itself. If the built-in speech-to-text isn’t good enough, you can try WhisperFlow. That’s what I use, and they offer a free plan as well. The second tip is create a spec. A spec is the set of rules for your project that

01:00 have to stay true, no matter how it gets built. [music] It’s how you deliver your understanding to AI. Here’s what Andrei has to say about it. >> I think people have to be in charge of this spec, this plan. I actually don’t even like the plan mode. I mean, obviously it’s very useful, but I think there’s something more general here where you have to work with your agent to design a spec that is very detailed, and maybe it’s uh basically the docs, and then get the agents to write them. And you’re in charge of the oversight and the the top-level categories, but the agents are doing a lot of the under the hood. >> [music] >> That might sound abstract, so let’s take a look at an example. Let’s say you ask

01:31 AI to build a booking form for your salon. It does a brilliant job, and the form might look perfect, but it lets people book on Sundays when you’re closed, and it lets two clients book the same slot. Something obvious to you might not be so obvious to an LLM. That’s why you need a spec. So, here are three practical steps to build one. First, uncover the goal. Make sure your agent knows exactly what you’re trying to achieve, what the end result looks like. Not build a booking form, but I want clients to be able to book themselves, so I stop answering the

02:01 phone all day. And don’t worry if you can’t say it off the top of your head. Just ask AI to interview you. Second, define the rules. Things like we’re closed on Sundays, only one client per slot, or anything else AI has no way of knowing. Because every assumption AI makes is a chance for it to drift from the final product you actually want. Finally, write it down. Now that you’ve uncovered the goal and the rules, ask AI to put them into a markdown file in your project. So next time it can just read it before it builds anything. Tip number three is give AI the right tools. Even

02:33 if your agent perfectly understands what you want, it still needs the right tools to complete the job. Let’s say you need to turn your script into a natural sounding voice over. Without a speech generation tool, your agent just won’t be able to do that. And when you choose one, you probably don’t want the result to sound like this. So, connect your agent to Fish Audio, which offers some of the most realistic AI voices out there. With Fish >> Audio, you can go with something natural and conversational. >> An expressive cartoon character. >> Or something a little more personal. >> You can make the delivery even more

03:04 natural with emotion tags. >> Make it sound really excited. Give it a little giggle. >> [laughter] >> Or keep it between you and me. >> Also, with Fish Audio’s API, you can give your agent the ability to generate these high-quality voice overs. For example, I have this folder with my script in a markdown file and a dot ENV file where I’ve saved my API key and the ID of the voice I want to use. I can now open this project in Code X and ask it to turn my script into a voice over using Fish Audio. The agent can now do

03:34 something it couldn’t do before. >> Open Claw is the most powerful AI tool ever created. >> So, check them out using the first link in the description. And shout out to Fish Audio for sponsoring today’s video. The next tip is give AI a finish line. AI models are incredibly smart. You don’t have to guide your agent every step of the way. Just tell it what you want to achieve and let it figure it out. Karpathy put it simply, “Don’t tell it what to do. Give it success criteria and watch it go.” Sometimes giving AI the exact steps can just restrict it.

04:05 You’re making it follow your approach even when there might be a better one. Tip number five is make AI challenge you. AI models can make your ideas sound incredibly convincing, but just because they agree with you doesn’t mean you’re right. Andre shared that he spent 4 hours working with an LLM to improve an argument for a blog post. And when he thought it was great, he asked it to argue the opposite. All of a sudden, it demolished the entire argument and convinced him that the opposite was in fact true. So, before you settle on an idea, ask AI to make the strongest

04:36 argument against it. You might find a weakness you missed while you were busy making your own argument sound better. [music] The next tip is build a knowledge base. When you research something with AI, you can end up with a lot of useful information scattered across different chats. Karpathy uses AI to organize his research into his own wiki. It’s basically a collection of notes in markdown files that your agent writes and updates for you. Let’s use this video as an example. I’ve spent a lot of time researching Karpathy’s interviews, tweets, and videos to find

05:07 his best tips on using AI. With a knowledge base, my agent could organize that research into markdown pages by topic with links back to the original sources. Then, when a new interview comes out, I could add the transcript and ask my agent to update those pages. So, the next time I’m working on a video about Andre Karpathy, I could ask it to start with the research we’ve already collected. If you want to try it out for yourself, I’ll leave the links to the original sources in the description down below. Finally, the bonus tip I promised at the beginning, and this one is my

05:37 favorite. It’s actually a quote Karpathy keeps bringing up. You can outsource your thinking, but you can’t outsource your understanding. And that reveals the foundation of how you should work with these agents. If you want to build a high-quality product, you need to actually understand what you’re building and stay involved in the process. And if something isn’t clear, just ask AI to explain it to you. The better you understand what you’re working on, the better you can guide your agent and judge its work. By the way, Fish Audio’s API is free until the end of November. So, if you want to try it yourself, check out the first link in the

06:08 description. So, if you found this video helpful, make sure you smash that like button and I’ll see you in the next one.