The AI Automators reviews Boris Cherny's claim that Anthropic removed most of Claude Code's system prompt for Opus 5 and routinely uses ablation to measure which instructions still improve behavior. The practical lesson is to test simpler configurations instead of carrying model-specific prompt fixes forever.
The video adds an important boundary: deny rules, destructive-command blocks, audit requirements and compliance checks are enforcement policies, not dead prompting. Long-running agents also need strong tests and feedback because small deviations can compound even when models can work autonomously for days.
Cherny's broader advice is to define the task, guardrails and exit criteria without over-prescribing implementation, then make the result verifiable. The analysis treats claims of prompt-injection immunity and multi-week autonomous runs cautiously because Anthropic has unusual resources and incentives. Course and community promotions are omitted.
Watch on YouTube
