Boris Cherny on prompting Opus 5 in Claude Code

Boris Cherny said Opus 5 scored 30% on ARC-AGI-3, agents now run autonomously for weeks, and Claude Code developers cut 80% of the system prompt.

Author: Michael Kokin ·

Boris Cherny (creator of Claude Code at Anthropic) gave a fresh interview to YC. Main focus — the recent Opus 5 release and how fundamentally our approach to working with models is changing. Spoiler: we're overcomplicating everything.

Key points from the conversation:

1. A breakthrough in reasoning: ARC-AGI-3
Opus 5 scored 30% on the ARC-AGI-3 benchmark. If you've tried these visual-logic puzzles yourself (I actually tested my own skills on them recently — it's brain-melting, I wrote about it in detail here), you know this is a huge leap for AI. The best previous results barely cracked 10-15%.

2. Total autonomy
The model can now autonomously handle a task not for minutes, but for days and even months. Cherny gave an example: he tasked an agent with rewriting a desktop app from Electron to Swift, running it on a Mac VM, taking screenshots, and comparing results pixel by pixel. The process has been running for over two weeks now, with the agent writing its own progress reports to Slack.

3. Prompt injection is nearly solved
Opus 5 is practically immune to jailbreaks. Thanks to mechanistic interpretability (Anthropic literally watches which neurons light up when someone attempts a jailbreak) and new classifiers, getting the model to execute a malicious instruction pulled from the web has become nearly impossible.

4. The product overhang effect
Developers stripped 80% of the system prompt out of Claude Code. Turns out the smarter the model gets, the more our rigid instructions and workarounds get in its way. Models can already do far more than we're asking of them, but our interfaces and habits are holding them back (unhobbling).

What this means for prompting

Talk to it more like a competent colleague, not a script.

You can watch the full video here.

!video