---
id: "lib-2026-07-27-boris-cherny-we-cut-80-of-claude-code-s-prompt"
type: "library"
title: "Boris Cherny: We Cut 80% of Claude Code’s Prompt"
lang: "EN"
org: "Unknown"
source_type: "Talk"
date: "2026-07-27"
week: "2026-W31"
url: "https://nosov.ai/library/lib-2026-07-27-boris-cherny-we-cut-80-of-claude-code-s-prompt/"
original_url: "https://www.youtube.com/watch?v=qyPCVqFUyDo"
tier: "B"
tier_name: "Useful, limited"
tier_reason: "First-hand practice from Boris Cherny, creator of Claude Code, on how Anthropic prunes its harness for each model; concrete and copyable, but every claim is the vendor's own and unmeasured in the talk."
summary_basis: "transcript"
original_minutes: "36"
tags: ["ai technology", "lecture"]
---

# Boris Cherny: We Cut 80% of Claude Code’s Prompt

Tier B (Useful, limited): First-hand practice from Boris Cherny, creator of Claude Code, on how Anthropic prunes its harness for each model; concrete and copyable, but every claim is the vendor's own and unmeasured in the talk.

Summary based on the full transcript.

## In short

A 36-minute Y Combinator interview (uploaded 27 July 2026) with Boris Cherny, the day after Opus 5 shipped. Each new model gets an "ablation": the team deletes the system prompt and tools, then adds lines back only where the product repeatedly fails. For Opus 5 they removed 80% of the prompt, and he says the model is slightly more capable with no prompt at all. His advice to builders: give the model a task a little harder than you think it can handle, describe guardrails and exit criteria instead of step-by-step instructions, and above all give it a way to verify its own work. He urges users to delete their CLAUDE.md, skills and hooks every six months and see what is still needed.

## Key numbers

- 80%: share of Claude Code's system prompt deleted for Opus 5
- 11 days: run time of a dynamic workflow that rewrote the Bun JavaScript runtime from Zig to Rust, with steering
- 20 or 30: daily maintenance routines Anthropic runs across its Claude Code code bases
- 30%: Opus 5 score on ARC-AGI 3 cited by the interviewer

## What it reports

- Claude Code's harness code is now mostly safety, permissions, static analysis and UI; much other code has been removed. An undocumented "simple" environment setting strips all prompts for testing.
- Evals live for one to three model generations before they saturate and need replacing.
- The Bun runtime was rewritten from Zig to Rust in 11 days using a dynamic workflow that orchestrates many agents against Bun's and Node.js's test suites; it is in production. He estimates more than a year of engineering by hand.
- A Swift rewrite of the Claude desktop app, verified by pixel comparison against the Electron version in a macOS virtual machine, had been running for 14-15 days at the time of the talk.
- Anthropic runs 20 or 30 daily routines that remove dead code, ship finished experiments, add or delete tests and merge near-duplicate abstractions across its apps.
- He claims Opus 5, with a prompt-injection classifier based on interpretability work and the auto-mode classifier, cannot be shown to be prompt-injectable in Anthropic's testing.
- He says coding is "solved" for his kind of work, but not for deep systems code, distributed systems or pixel-level UI checks.

## How they got it

Personal and team practice inside Anthropic, described in an on-stage interview. No metrics, error rates or costs are shown for the long-running jobs or routines.

## Limits

- Vendor speaker at a launch-week event; claims about prompt injection and multi-week runs are unverified here.
- Results come from a team with unusual access to compute and to the model's developers.
- Live-event transcript with audience asides.

## Why it matters for companies

The transferable practice is the ablation habit: when the model changes, remove instructions and add back only what failures justify. Most enterprise agents carry prompts and rules written for older, weaker models, and these cost tokens and can lower quality. The stronger message is about verification: long autonomous jobs worked here because a test suite or a pixel comparison told the agent when it was done. Teams planning agent work on HR, finance or document processes should first build that check; without it, extended runs have no stopping rule.

Original: https://www.youtube.com/watch?v=qyPCVqFUyDo