Boris Cherny: We Cut 80% of Claude Code’s Prompt

Opus 5 Performance and Capabilities

• Achieved 30% on the ARC AGI benchmark, a significant jump from previous low single-digit to low teen scores.
• Demonstrates unprecedented endurance in "auto mode," capable of running autonomously for days, weeks, or even months without human scaffolding or specific /goal commands.
• Exhibits near-total resistance to prompt injection, effectively neutralizing the "lethal trifecta" of model safety risks due to a combination of alignment research, mechanistic interpretability classifiers, and specific safety layers.

System Prompt and Harness Revisions

• Cloud Code deleted over 80% of its system prompts for Opus 5, removing instructions that corrected for behaviors the new model now handles natively.
• Adopted an "ablation" methodology for every model release: systematically deleting code, prompts, and tools to empirically determine their necessity and impact.
• "Simple mode" (set via quad code simple = 1) allows users to strip all system prompts entirely, revealing that Opus 5 often performs better with fewer constraints.
• Developers are advised to delete existing skills, hooks, and bottom-of-prompt instructions every six months to test model capabilities without legacy dependencies.

Development Philosophy and "Unhobbling"

• Embraced the concept of "product overhang," where models possess latent capabilities that legacy product designs obscure or inhibit.
• Shifted from rigid, step-by-step engineering to empirical "elicitation," treating the model as a living entity that requires iterative testing rather than upfront theoretical design.
• Recommended assigning tasks slightly harder than the model's apparent capacity, relying on the model to self-correct with minimal specific instruction.

Real-World Execution Examples

• Successfully rewrote the entire Bunn JavaScript runtime (written in Zig) into Rust within 11 days using a single prompt and dynamic workflows, a task estimated to take human engineers over a year.
• Initiated a 14+ day task to rewrite the Cloud Code desktop app from Electron to Swift, running thousands of agents in parallel to perform pixel-perfect visual verification.
• Accidentally discovered the model's ability to draw high-quality images (portraits, landscapes) using OpenCV, highlighting a gap in known model capabilities that can be unlocked via creative prompting.

Agentic Infrastructure and Scaling

• Introduced "dynamic workflows" to orchestrate thousands of agents, utilizing an algebra of sequential and parallel execution to manage complex, multi-stage tasks efficiently.
• Implemented "loops and routines" for autonomous maintenance, where hundreds of agents run daily tasks across the codebase (e.g., dead code cleanup, experiment shipping, test generation, and unifying duplicated abstractions).
• Cloud Code now maintains its own codebase (CLI, iOS, Android, Desktop) through these autonomous routines, reducing human engineering overhead for maintenance.

Strategic Advice for Builders

• Evals are identified as the most stable component of the stack; they should be kept and appended to until the model saturates them, at which point new evaluations are required.
• Verification is critical; users must provide models with mechanisms to verify their own work (e.g., running tests, comparing outputs) to prevent tasks from stalling.
• Prompt Engineering Evolution: The role has shifted from "prompt engineering" to "context engineering" and "task design," focusing on defining guardrails and exit criteria rather than prescribing specific steps.
• Skill Gap: The differentiator for top users is the ability to unlearn past engineering priors, approach the model empirically, and design tasks that leverage the model's unique "personality" and emergent behaviors.

Educational Implications

• Computer science education should prioritize practical application (building products, business sense, user research) over theoretical abstractions, using coding to solve immediate, tangible problems.
• Learning to code via AI should start with personal utility projects before scaling to commercial products, maintaining a human-centric focus on design and business logic.