
The AI Automators
Boris Cherny Just Told Us To Delete Our CLAUDE.md Files
Summarised with Bite · 11 min read
This video dissects a recent Boris Cherny talk about Claude Code, separating the genuinely useful lessons from the marketing heat. The big takeaway is not simply "delete your CLAUDE.md," but learn when to remove stale prompt scaffolding, when to keep hard enforcement, and how modern coding agents only become powerful when they can verify their own work over long runs.
0:00 – 4:07
Delete the prompt, but not the seatbelt
The video opens with a provocative claim: the creator of Claude Code is basically telling people to trash their carefully crafted CLAUDE.md files. That sounds reckless until you hear the reasoning. Boris Cherny argues that newer frontier models can now sustain long-running work for days or even weeks, especially when paired with Claude Code's auto mode, and that they need far less external orchestration than before. The host agrees there is a real trend here, models are getting better at staying on task for longer, but immediately adds the missing context: Boris works at Anthropic and has a nearly unlimited token budget, so his baseline is not the same as an ordinary user paying for usage. From there, the talk moves into a claim that the host clearly finds overstated: Boris says Claude is no longer meaningfully prompt injectable. The explanation is that Anthropic now relies on three layers, a more aligned base model, a prompt-injection classifier over traffic, and an auto-mode classifier to block dangerous actions. The host's reaction is blunt. Claude may be getting more resistant, but saying prompt injection is effectively solved is a bridge too far. The more practical part comes next. Boris says that when Opus 5 shipped, they deleted more than 80% of Claude Code's system prompt. That is the surprising angle of the whole video: progress in the model often means deleting product logic, not adding more. The host demonstrates two concrete ways users can test this idea themselves. One is manually overriding the system prompt with a command and making Claude respond in a silly Klingon-English mix, which proves the output is still steerable. The other is setting an environment variable, Claude code simple equals one, which strips back system prompting and tool prompting so the team can test what is actually necessary. That leads to the real lesson, ablation. Anthropic repeatedly removes prompts and tooling, then adds them back piece by piece to measure what actually helps. The host agrees this is powerful, but adds an important warning that Boris's slogan, "every 6 months, delete your CLAUDE.md file," needs translation. Delete dead weight, yes. Delete your destructive-command blockers, deny rules, compliance hooks, or audit constraints, no. Those are not coaching hints for the model. They are enforcement policies. In other words, remove the training wheels if the rider has improved, but do not remove the brakes from the bike.
2 more sections in the app
- 4:07 – 8:45When less scaffolding wins, and when it absolutely does not
- 8:45 – 15:26The real bottleneck is verification, not clever prompting




