Boris Cherny Just Told Us To Delete Our CLAUDE.md Files

👉 Access our AI Architects course & join hundreds of serious AI builders in our community: https://www.theaiautomators.com/?utm_source=youtube&utm_medium=video&utm_campaign=tutorial&utm_content=boris-guidance 🔗 The talk Boris Cherny's full interview: https://www.youtube.com/watch?v=qyPCVqFUyDo 🔗 Changing the Claude Code system prompt yourself Claude Code CLI reference (the --system-prompt flag): https://code.claude.com/docs/en/cli-reference Claude Code environment variables reference: https://code.claude.com/docs/en/env-vars 🔗 The Zig to Rust rewrite Bun repo: https://github.com/oven-sh/bun Bun docs: https://bun.com/docs/ Claude Tag, Anthropic's Slack agent: https://techcrunch.com/2026/06/23/anthropics-claude-tag-is-learning-your-company-one-slack-message-at-a-time/ 🔗 Our related videos Our video on coding harnesses (the ablation study): https://www.youtube.com/watch?v=QpceyQQwC_E Our video on loop engineering: https://www.youtube.com/watch?v=5WB5bcGNib8 The creator of Claude Code just told us to delete our CLAUDE.md files. He also explained how he gets Claude Code to run autonomously for weeks at a time. And he claimed the model is no longer prompt injectable. Boris Cherny gave this talk the day after Opus 5 launched, and there is genuinely good advice in it. When Opus 5 shipped, his team deleted over 80% of the Claude Code system prompt, because behaviour tuned for one model often does not transfer to the next one at all. He wants you to run the same ablation on your own setup. Stop over-specifying your prompts, define the task, the guardrails and the exit criteria, then let the model work. Give it problems that are slightly too hard, and make it possible for Claude to verify its own work. That last one is the part most people get wrong. But some of it needs a pinch of salt, and that is where this video goes. Three layers of prompt injection defence is real progress. Not being demonstrably prompt injectable is not the same thing as not being injectable, and I would not build on that assumption. Delete your CLAUDE.md every six months is good hygiene for instructions, but hooks that block destructive commands, deny rules and compliance constraints are enforcement policies, not dead weight the model will outgrow, so those stay. And the famous Bun rewrite, a whole codebase from Zig to Rust in 11 days of runtime, came with steering throughout and solid test suites underneath it. Then there is the bill. At the time of the interview he had a Claude Code run going for a little over two weeks, 14 or 15 days, and it was still running. He has a practically unlimited token budget. An experiment posted online suggested a $200 a month Max 20x plan can return as much as $8,000 in equivalent API credits, so even a subsidised plan gets you nowhere near that volume. On API pricing, a multi-week run on frontier models is tens or hundreds of thousands of dollars, and long running agents drift, so by day nine or ten you can be a long way off course before anyone notices. He is clearly a very smart guy and an extreme power user, and I come out of his interviews with real experiments to try. He also works at Anthropic, and Claude Code is his creation so he has huge incentives to push it to the max. In this video I go through the claims and give you my verdict on each, along with the two levers you can pull today to strip back the prompting yourself and see what your model can actually do without you constraining it. Timestamps: 0:00 The three claims 0:25 What Opus 5 actually changed 1:10 "No longer prompt injectable" 1:52 80% of the system prompt, deleted 2:15 Two levers you can try yourself 3:20 What I would keep, whatever he says 4:10 Overhang, hobbling, and underhang 6:20 Stop over-specifying 7:39 Zig to Rust in 11 days 8:45 Verification is the real skill 9:40 The LinkedIn influencer line 10:30 What a two week run actually costs 11:55 Workflows vs loops and routines 13:10 Coding is solved, with an asterisk 14:15 The verdict