Claude Code Creator Boris Cherny Urges Developers to Delete 80% of System Prompts

Every six months, delete your Claude.md, delete your skills, delete your hooks.

This is Boris Cherny, the creator of Claude Code, offering advice to product users.

In the newly released YC interview video on July 28, “Boris Cherny: We Cut 80% of Claude Code’s Prompt,” Boris enthusiastically urges all AI product builders to boldly delete elements from their new products—aggressively removing system prompts, tools, and harness code.

△

You should delete the entire system prompt and then add it back line by line to see what effect each line has.

The concept behind this idea is an ablation study, which runs throughout Boris Cherny’s latest interview, aiming to remove, replace, or disable a module while holding all other conditions constant, in order to compare changes in performance, stability, efficiency, or cost.

In the interview, Boris proudly stated: "For Opus 5, we genuinely recommend deleting all of these, as the model no longer needs them."

Despite the recent turmoil... users can also say, "Honestly, I'd recommend deleting Claude because we no longer need it." (x)

In addition to "bold deletion," Boris also shared deeper insights in the interview regarding his thoughts on the value of product design, model usage, and learning to program—

The views in this article are drawn from an interview video. Key takeaways include: 1. Boris’s product iteration strategy: Predict less, test more 2. “Product slack”: Model capabilities will always outpace product boundaries over time 3. “Unshackling”: Challenge models with harder tasks and allow them to work independently for longer periods 4. How do top Claude users utilize AI? 5. Three recommendations for programming learners

Today, the code in the Claude Code harness has been reduced almost entirely to security, permissions, and static analysis components.

On July 24, Anthropic released updated guidelines for Claude 5's context engineering, significantly streamlining the system prompt for Claude Code—removing over 80% of the original instructions—for new models such as Opus 5 and Fable 5.

For details, refer to the QbitAI article: "Claude Code Deletes 80% of Prompts, Opus 5 Reverses It."

Regarding this change, Boris shared his product iteration strategy in the interview: Don’t try to guess what instructions the model needs, because you’ll never get it right. All you can do is delete line by line, test, and identify where the model repeatedly gets stuck.

Think of the model as a living organism, something more organic. Each generation behaves differently and has slightly different characteristics—you need to spend time understanding it and adjust the harness accordingly.

△

Therefore, for Boris, this is more “empirical”—it requires a scientific approach: try without preconceptions, observe the results, iterate, and repeat.

In this ever-evolving landscape, Eval may also fail to remain stable. Although it is indeed more durable than harness and prompt, models are advancing so rapidly that a set of evaluations is often quickly scored perfectly. Therefore, it’s necessary to observe where the model consistently struggles and then design new Evals.

In the interview, Boris shared a concept called "Product Overhang," which he believes is a helpful way of thinking about product development.

Overhang means surplus.

Large models advance in discontinuous leaps, while product integration progresses at a continuous, incremental pace. This causes the capabilities of the models to consistently exceed the boundaries of what current products can deliver.

Boris gave an example: When Sonnet 3.5 was first released at the end of 2024, the model was already capable of writing entire files of code in one go. At that time, early versions of programming products like Copilot and Cursor were still focused on minor tasks like code completion.

Claude Code, with full terminal access, somewhat bridges this gap. This is also Boris’s second concept: “Unhobbling”—removing restrictions.

△

He shared an internal case from Anthropic: someone attempted to connect Opus 5 with OpenCV (the world’s largest open-source computer vision library) and discovered that the model could autonomously generate portraits and animal landscapes—something it had never been trained to do before.

This is "model elicitation"—unlocking capabilities that the model already possesses but has not previously been activated, through the design of prompts, context, tools, or product interfaces, without altering the model's weights.

There may still be questions here: How should we attribute the results? Did the model simply demonstrate abilities it already possessed, or did it learn new skills due to the “scaffolding” design?

But this may not matter at all; regardless, Boris has no doubt that there are enormous business opportunities here:

I’m not saying every startup can seize it, but I know people are thinking about these issues—there’s truly enormous potential here to elicit remarkable, interesting, and commercially valuable behaviors from models.

To this end, Boris proposed three of his own model "unbinding" methods.

First, give the model tasks that are harder than you think. Clearly define the goal, boundaries, and exit conditions, then let go.

Second, conduct more experiments. Allow the model to make playful attempts without clear commercial goals—“give yourself the freedom to experiment with the model and be creative.”

Third, let the model verify its own results. The focus today is no longer on "prompt engineering"—the question is, "Once you assign Claude a very difficult task, how do you get it to verify its own work during the process?"

Boris believes the third point may be the one most poorly executed today, because if a model cannot verify its own tasks, it cannot operate independently for extended periods.

△

His own example might inspire everyone:

Boris: "Okay, what I want you to do is rewrite the Electron app in Swift. Run the Electron app inside a Mac virtual machine, take screenshots, and then compare them pixel by pixel against the Swift version—don’t stop until it’s done." Host: "That’s your prompt?" Boris: "That’s my prompt." Host: "How long has it been running?" Boris: "It’s still running." Host: "When did you start?" Boris: "It’s been running for over two weeks—about 14 or 15 days... Claude also decided to livestream it; internally, it created a Slack channel and posts a progress screenshot every few minutes."

At the end of the interview, the host posed the question:

So, Boris, how can we use Claude as well as you do?

Boris said that the most important thing is to ignore what influencers on LinkedIn say and avoid scrolling through Twitter.

Regarding the use of AI, "everyone is looking for that kind of 'secret trick.' But it doesn't exist. There's no such thing."

He suggested approaching the model empirically, setting aside past experiences with older models and computer science theories learned in class, and instead directly observing where the model struggles before making targeted adjustments.

So it’s no longer a theoretical science—it has become an empirical one. I believe those who are especially good at letting go of their preconceptions, abandoning the mindset of “it didn’t work before,” and are willing to try again will be extremely successful.

More importantly, maintain a mindset of letting go of your desire to control the model—treat it like a colleague: avoid overly specific instructions, don’t demand excessive detail, and don’t try to make it perform tasks exactly the way you would. Because “that’s not how models work.”

For those still learning to code, Boris urges them not to focus solely on pure computer science theory, but to learn how to apply it. For example, his own initial motivation for learning programming was to cheat on math exams.

Usually, it’s about entrepreneurship, building products, cultivating your own sense of design and business, learning how to do data science, and learning how to talk to users... Only when you combine these with computer science and engineering do they become truly valuable.

In short, "first build what you want, then upgrade to build what others want."

This article is from the WeChat official account "Quantum Bit" (ID: QbitAI), authored by: Focused on frontier technology.