Claude Code's creator says better prompts need just 3 things, so I tested his approach

We've successfully reached the stage where every other person has an opinion on how you should be prompting AI. In the same week, you'll hear one person tell you to feed as much context as you possibly can into a prompt, be vague, and let the LLM figure out the details on its own. Then, you'll come across someone else telling you to do the exact opposite: keep your prompts short, be ridiculously specific about what you want, and leave absolutely nothing up to the AI's interpretation.

Sure enough, both sides will swear their approach is the secret to getting better results. This is why I think the best tips come from the people actually building these models and the tools powered by them. After all, who better to tell you how to get the most out of an AI tool than someone who played a role in creating it?

That's exactly why a recent post from Boris Cherny, the creator of Claude Code, caught my attention. Rather than sharing some shiny new, elaborate prompting framework or a ridiculously long template you need to follow, he says a good prompt just needs three things.

Boris Cherny says there's no secret to prompting

Three tips, no magic formula

Cherny frequently shares his Claude Code setup tips over on his X account, and a post he shared on October 6, 2026, caught quite a bit of attention. In the post, he shared an interactive website he'd built using Claude to accompany an episode of the Acquired podcast about Home Depot. Interestingly, it wasn't the website itself that got people talking. Instead, it was the prompt he'd used to create it.

Once he shared the prompt, users were surprised by just how simple and conversational it was. There were no elaborate instructions, complicated prompting frameworks, or special tricks involved. He was essentially talking to Claude the way you'd talk to a coworker, which led quite a few people to realize they'd been prompting their AI tools in a similar way all along.

The reactions eventually prompted Cherny to elaborate on his approach to prompting, and that's when he explained that there's really no secret to it. According to him, a good prompt comes down to just three things.

First up, your prompt should clearly communicate what you want it to do. This is fairly self-explanatory, but it essentially means being clear about what you're trying to accomplish rather than leaving the AI to guess what you have in mind. You don't necessarily need to write an entire paragraph explaining every little detail, but the model should have a good understanding of what you're asking it to do and what you expect the final result to look like.

Something Anthropic's been pushing a lot lately is giving users more control over how much time and computational effort a model spends working through a task, which is referred to as effort. The idea is pretty straightforward: not every task requires the same amount of thinking. You wouldn't necessarily want Claude to spend several minutes reasoning through a simple question, but for something considerably more complicated, giving it more room to think could make a difference. That's essentially what Cherny's second tip boils down to: communicate exactly how much effort you want the AI to put into the task.

Finally, another tip that Cherny has shared in practically all of his Claude Code setup tip articles is giving the tool a way to verify its own work. Rather than simply trusting that the AI has done what you asked it to do correctly, give it a way to check its output. For example, if you're asking Claude Code to build an application, you could ask it to run tests to make sure everything works as intended. The idea is to give the model a clear way to determine whether it's actually accomplished what you asked it to do, rather than just assuming it has.

Cherny's own prompt is an almost perfect example of his advice

He practices (almost) what he preaches

While I could demonstrate Cherny's advice with a prompt I've written myself, I think there's no better example than the very prompt that started this entire conversation in the first place. After all, if we're taking prompting advice from the creator of Claude Code, we might as well look at how he puts his own advice into practice.

The latest acquired Home Depot episode is really amazing. I want you to make an artifact that narrates the entire episode. Make it beautiful and interactive and focus on communicating the nugggets of business wisdom throughout the episode. Make it really good, use lots of tokens and iterate till you're proud of it. Maybe add watercolor illustrations and historical photos throughout, aim for New Yorker or nytimes data viz vibes but better. I think you can use OpenCV for water color images

Let's start with the first tip: clearly communicating what you want the AI to do. Cherny begins his prompt by asking Claude to create an artifact that narrates the entire Home Depot episode. He then gives it a few more details about what he has in mind. He wants the artifact to be beautiful, interactive, and focused on communicating the business lessons from the episode. He even gives Claude some creative direction, suggesting watercolor illustrations, historical photographs, and a visual style inspired by The New Yorker or The New York Times.

What's interesting is that he doesn't tell Claude exactly how to build the artifact. He explains what he wants the final result to look like and leaves the finer details up to the model. The second tip is even easier to spot. Cherny explicitly tells Claude to "use lots of tokens" and keep iterating until it's proud of the result. He's essentially telling the model that he wants it to put considerable effort into the task rather than rushing through it.

Now, the third tip is where things get a little interesting. While Cherny does ask Claude to keep iterating until it's satisfied with the result, he doesn't actually give it a concrete way to verify its work. There are no instructions to run tests or check the final artifact against a specific set of requirements. Instead, he's relying on Claude to evaluate and improve the quality of its own output.

So, while his prompt clearly demonstrates the first two principles, the third is a little less explicit. I think that's especially worth pointing out, because it shows that even Cherny doesn't necessarily treat these three tips as a rigid formula that every prompt needs to follow.

After testing his approach out over the past few days, I've realized that there's a lot of truth to what Cherny is saying. Rather than spending ages trying to craft the perfect prompt or following some elaborate framework I came across online, I've been focusing on clearly communicating what I want, how much effort I expect the AI to put in, and how it should check its work. So far, I've found that this simple approach works surprisingly well, regardless of which AI model I'm using.

The problem with prompting advice is that it doesn't age well

Prompting advice has a surprisingly short shelf life

To be completely fair, when I said that the best prompting tips come from the people actually building these AI models and tools, I didn't mean that everyone else's advice is necessarily wrong. Sure, the LinkedIn prompt engineering gurus might not have all that many credentials to back up their advice, but that doesn't automatically make their tips any less useful. The problem is that prompting advice has a surprisingly short shelf life, and something that works brilliantly today might not work nearly as well a few months down the line.

If you've been following the pace of these AI companies since the day OpenAI dropped its first model publicly, you might've noticed that they're showing no signs of slowing down. A new model seems to drop practically every other week (or at least that's what it feels like), and the entire point of these releases is to improve upon what came before them. Naturally, that also means the way we interact with these models is constantly changing.

What worked brilliantly with one model might not necessarily work as well with the next, and prompting advice that was considered gospel just a few months ago might already be outdated. To add to that, every AI provider's LLMs are different. A prompting technique that works wonders with Claude might not necessarily produce the same results with ChatGPT or Gemini, even if you're asking all three models to perform the exact same task.

So, between the constant stream of new model releases and the differences between models from competing AI companies, it's no surprise that everyone's constantly trying to figure out the next big prompting trick.

These three tips aren't just for Claude

While Cherny's advice comes from his experience building and using Claude Code, none of these tips are exclusive to Claude. Whether you're using ChatGPT, Gemini, or a local LLM, clearly communicating what you want, how much effort the task requires, and how the model should verify its work can help you get better results.

Ultimately, these aren't complicated prompting tricks that depend on a specific model or tool, but rather simple principles that can be applied to just about any AI assistant!