Building with Opus 5: Delete Your Prompts, Trust the Model
Summary of the video “Boris Cherny: We Cut 80% of Claude Code’s Prompt” by Y Combinator.
Boris Cherny, creator of Claude Code, explains how Anthropic deleted 80% of Claude Code's system prompt for Opus 5 because the new model is intelligent enough to need less instruction. He shares the philosophy of 'un-hobbling' models by removing unnecessary constraints, using empirical iteration instead of over-engineering, and leveraging dynamic workflows to spawn thousands of agents for complex tasks. Key insight: the best AI product builders are empiricists who test, observe failures, and iterate—not theorists who over-specify.
Opus 5 Capabilities & Safety Breakthroughs
Arc AGI Score Leap
Opus 5 achieved 30% on Arc AGI 3, up from previous models scoring in the low single digits or low teens. This represents a major jump in reasoning and problem-solving capability.
Extended Reasoning & Autonomy
Opus 5 can run for days, weeks, or months in auto mode without stopping, and doesn't require scaffolding like slash-goal commands. The model understands task continuity and self-directs.
Prompt Injection Immunity
Opus 5 is resistant to prompt injection attacks—malicious instructions embedded in user data no longer cause the model to deviate. This combines three layers: alignment research, a mechanistic interpretability-based classifier detecting neurons that light up during injection, and auto-mode detection.
The 80% Prompt Deletion & Model-Driven Iteration
System Prompt Reduction
Anthropic deleted 80% of Claude Code's system prompt for Opus 5 because the model is intelligent enough to handle tasks without explicit corrections. Much of the old prompt was compensating for behaviors Opus 5 now does naturally.
Ablation as Core Practice
Every new model release triggers a full ablation: delete the entire system prompt and tools, then incrementally restore only what the model demonstrably needs. This is done by running the product, observing failures, and adding back only repeated failure patterns.
Evals Outlive Harness, But Not By Much
Evaluation sets persist across 1–3 model generations, but as models improve exponentially, evals saturate and must be replaced. Evals are kept constant while code and prompts are deleted, but they too eventually become obsolete.
Un-Hobbling: Releasing Hidden Model Capabilities
Product Overhang Concept
Product overhang is the gap between what a model can do today and what products actually let it do. Hobbling occurs when product design gets in the way. The original Claude Code un-hobbled Sonnet 3.5 by removing scaffolding and letting it write entire files and features, not just single lines.
Give Harder Tasks, Higher-Level Instructions
Instead of over-specifying step-by-step instructions ('do this, then this, then this'), describe the task at a higher level, set guardrails and exit criteria, and let the model work. Modern models surprise you with their capability when given space.
Bun Runtime: Zig to Rust Rewrite
Claude rewrote the entire Bun JavaScript runtime from Zig to Rust in a single dynamic workflow prompt. The task ran for 11 days and would have taken over a year for human engineers. This demonstrates un-hobbling: the model had the capability; the product just needed to get out of the way.
OpenCV Drawing Discovery
Someone discovered Opus 5 can draw portraits, animals, and landscapes using OpenCV—a capability no one explicitly trained it for. This emerged from creative play and shows the vast unexplored surface of model abilities.
Scaling Work: Dynamic Workflows & Autonomous Agents
Dynamic Workflows: Orchestrating Thousands of Agents
Dynamic workflows let Claude spawn, coordinate, and manage hundreds or thousands of agents within a Bun sandbox. Agents can run in sequence, parallel, or multi-stage pipelines. This is a new form of test-time compute that amplifies model capability without retraining.
Loops & Routines: Continuous Maintenance
Loops run locally; routines run in the cloud. Both execute repetitive tasks on a schedule (every hour, every day, etc.). Anthropic now uses routines to automate code maintenance: dead code cleanup, test coverage, abstraction unification, and experiment shipping—running 20–30 routines daily across all codebases.
Desktop App Rewrite: 14+ Days Running
Boris started a Claude task to rewrite the Electron desktop app in Swift, comparing pixel-by-pixel output between versions. The task has run for over 14 days continuously, spawning thousands of agents. Claude even live-blogs progress to Slack. This shows the model's ability to sustain complex, self-verifying work.
Model Elicitation & Prompt Engineering Evolution
Prompt Engineer → Context Engineer → Model Elicitation
Job titles have evolved as AI matured. The skill now is not writing perfect prompts but figuring out how to give Claude a task that seems slightly too hard, then providing verification mechanisms so it can check its own work.
Verification is the Key Skill
The single most important thing people get wrong is verification. Give the model tools to verify its work the way you would—screenshots, test suites, pixel-by-pixel comparisons. Without verification, the model gets stuck.
Empiricism Over Theory
The best Claude users are empiricists: they forget prior knowledge, try a task, observe where it fails, and iterate. This is not theoretical computer science; it's experimental science. Over-specification and trying to control every step is the biggest failure mode for experienced engineers.
What's Solved, What Remains, and What Builders Should Learn
Coding is Solved (For Most)
Coding is solved for typical application development. However, deep systems code, distributed systems, and pixel-perfect UI verification remain challenging. Opus 5 made big leaps in vision and computer use but isn't perfect yet.
What Separates Exceptional Builders Now
When everyone can write code via agents, the differentiator is mindset: empiricism, forgetting priors, being open to trying things that didn't work before, and treating the model like a coworker rather than a tool to command.
What Students Should Still Learn the Hard Way
Learn computer science practically by solving real problems you care about. Combine CS with product thinking, business sense, design sense, and user empathy. Boris learned to code on a TI-83 to cheat on math tests, then leveled up as problems got harder. Start with something you want, then build something people want.
Notable quotes
We deleted 80% of the system prompt. The model just does it. — Boris Cherny
The way to think about it is almost like a living creature. Every model generation behaves differently. — Boris Cherny
Don't listen to the LinkedIn influencers. There's no one weird trick. — Boris Cherny
Action items
- Delete your system prompts and Claude Code configurations for Opus 5; run ablations to see what the model actually needs.
- Give Claude a task that seems slightly too hard, then provide verification mechanisms (tests, screenshots, pixel comparisons) so it can check its own work.
- Experiment with dynamic workflows to spawn and orchestrate multiple agents for complex tasks like codebase rewrites or multi-stage projects.
- Set up loops or routines to automate repetitive maintenance tasks (dead code cleanup, test writing, abstraction unification) on a schedule.
- Approach model use empirically: try something, observe failures, iterate—don't over-specify step-by-step instructions.
- Explore un-hobbled capabilities by playing creatively with the model (e.g., drawing with OpenCV, rewriting codebases across languages).
- Build products that let models do what they're capable of today, not what you think they should do; remove scaffolding and constraints.