By playing, you agree to uListen's Terms and Privacy Policy and YouTube's Terms of Service.
Summaries, chapters, transcripts, and FAQs on this page are made by uListen, not YouTube.
At a glance
WHAT IT’S REALLY ABOUT
Why Claude Code deleted most prompts and scaled long-running agents
- Boris Cherny says Opus 5 differs from prior models by sustaining extremely long task runs (especially with Auto Mode) and showing a new level of resistance to prompt injection.
- He explains Claude Code’s team deleted over 80% of its system prompt because Opus 5 no longer needs many legacy behavioral “fixes,” and their process is to ablate prompts and add back only what repeated failures justify.
- He argues effective model use has shifted away from ornate prompt engineering toward giving harder tasks plus strong verification mechanisms so the model can check its own work as it iterates.
- He introduces “product overhang” and “unhobbling” as the opportunity to redesign products so models can express capabilities they already have, rather than being constrained by outdated scaffolding or overly specific instructions.
- He describes mechanisms for scaling agentic work—dynamic workflows and cloud/local routines—and gives examples of using them for large rewrites and ongoing codebase maintenance.
IDEAS WORTH REMEMBERING
5 ideasNewer models can make large parts of old system prompts redundant.
He says Claude Code’s system prompt and tool prompts had accumulated workarounds for older model shortcomings, but Opus 5 now “just does it,” so much of that text became unnecessary and even slightly harmful to intelligence in ablation tests.
Rebuild prompts empirically: delete first, add back only after repeat failures.
He describes their process as deleting the whole system prompt and reintroducing lines only when they observe repeated, specific failures—because every instruction is re-read on every run and should “earn its place.”
Prompt injection defense works best as multiple stacked layers, not just prompting.
He attributes improved robustness to prompt injection to three layers: a well-aligned model, a prompt injection classifier informed by mechanistic interpretability signals, and an Auto Mode classifier; with these, they “cannot demonstrate prompt injection anymore.”
Modern models perform better with higher-level direction plus clear guardrails than micromanaged steps.
He argues many users “hobble” models by overspecifying steps; instead, they should describe the task, guardrails, and exit criteria, then let the model work, checking in later—an approach he says didn’t work as well six months prior but does now.
“Unhobbling” is often a product design problem: remove constraints that block model capability.
He frames “product overhang” as capabilities the model already has but products fail to elicit, and says early Claude Code aimed to remove scaffolding so models could write at the file/feature level rather than being constrained to autocomplete or read-only chat.
WORDS WORTH SAVING
5 quotesFor 5, one example of something it does that I think no other model has done is it runs for a very long period of time, and especially when you combine Opus 5 with Auto Mode, it's just, like, incredible. Like, it can go for days, weeks, months at a time. It just won't stop.
— Boris Cherny
So yeah, we deleted eighty percent of the system prompt.
— Boris Cherny
100%. Yeah, and, and for people that aren't building agentic products, but you're using Claude Code, every six months, delete your Claude MD, delete your skills, delete your hooks. See what the model does, and it might surprise you.
— Boris Cherny
The way to think about it is almost like a, like a living creature, like a something more organic. It's a thing where every model generation, it behaves differently. It has a slightly different personality, and you have to take the time to get to know it and then adjust the harness based on that.
— Boris Cherny
The skill nowadays is less about prompt engineering and more about figuring out how do you give Claude a hard task that seems a little bit too hard. And then how do you make it possible for Claude to verify its work along the way?
— Boris Cherny
SPEAKERS
Diana Hu
hostHost/interviewer for Y Combinator conversations.
Boris Cherny
guestCreator of Claude Code (Anthropic), speaking about building and iterating on its prompt/harness and agentic workflows.
High quality AI-generated summary created from speaker-labeled transcript.
