We Cut 80% of Claude Code's Prompt: Boris Cherny on Radical Simplicity

When a new model ships, delete your system prompt, CLAUDE.md, skills, and hooks first. Most of the rules you stacked up for older models are dead weight now, and deleting them often makes the model smarter.

Don't refill deleted rules by guessing. Run the product for real, and add a line back only when the model stumbles over the same thing repeatedly. Ablation is an eval where you delete things to measure each line's impact.

Treat the model like a living creature, like a coworker. Each generation has a different personality, so don't port over what worked on the last model — try, observe the result, and adjust empirically.

Give a task that looks slightly too hard. Instead of dictating steps one through four, describe the task, the guardrails, and the exit criteria, then let the model cook. Over-specification hobbles modern models.

Verification is the single most important lever. Hand the model a hard task, but always give it a way to check its own work — then it won't get stuck and will run for days or weeks on its own.

Chase the product overhang. Keep throwing the newest model at old problems, and experiment creatively even when there's no obvious commercial payoff. The model already holds abilities no one has elicited yet.

Orchestrate thousands of agents in sequence and parallel with dynamic workflows, and hand repetitive work to local loops and cloud routines so the codebase maintains itself.

Evals outlive code and prompts, but only by a generation or two. Keep appending, expect saturation, and treat throwing an eval away and building a new one as part of the process.