Head of Claude Code: What happens after coding is solved | Boris Cherny
Boris Turney, Head of Claude Code at Anthropic, discusses how AI has transformed software engineering, enabling 100% AI-written code and increased productivity. He explores the future where AI generates ideas, branches beyond coding, and the societal shifts agentic AI will bring, offering advice for adapting.
Deep Dive Analysis
17 Topic Outline
Boris's Brief Departure from Anthropic and Return
Claude Code's One-Year Impact on Software Engineering
The Origin Story and Evolution of Claude Code
The Exponential Pace of AI in Software Development
Boris's 100% AI-Written Coding Workflow
The Next Frontier: AI Generating Ideas and Beyond Coding
Principles for Building AI Products and Teams
The Printing Press Analogy for AI's Impact
Future Roles Most Impacted by Agentic AI
Advice for Succeeding in the AI Era
The Principle of Latent Demand in Product Development
The Rapid Development of Cowork in 10 Days
Anthropic's Three Layers of AI Safety
Anxiety and Productivity with AI Agents
Pro Tips for Using Claude Code Effectively
Boris's Post-AGI Plans
Reflections on Claude Code's Growth and Future
5 Key Concepts
Latent Demand
Latent demand is the idea that if a product can be 'misused' or adapted by people in ways it wasn't designed for, it indicates a strong underlying need. This helps product builders understand where to take the product next by observing user behavior and building specific solutions for those emerging use cases.
Agentic AI
Agentic AI refers to an LLM that is capable of using tools and acting in the world, not just conversing. This means it can interact with systems like Gmail, Slack, or a computer's command line to perform tasks, moving beyond simple chatbot functionality.
The Bitter Lesson
This principle, articulated by Rich Sutton, suggests that a more general model will almost always outperform a more specific model over the long term. In AI product development, this implies betting on larger, more general models rather than over-optimizing or fine-tuning smaller, specific ones.
Mechanistic Interpretability
This is a field of study that aims to understand what is happening inside the 'neurons' of AI models, similar to studying a human or animal brain. It helps researchers trace how particular concepts are encoded, how the model plans, and how it thinks ahead, contributing to AI safety and alignment.
Multi-Quadding
Multi-quadding refers to the practice of running multiple Claude Code or Cowork sessions in parallel. This allows users to kick off several tasks simultaneously and let the AI agents work independently, significantly boosting productivity.
8 Questions Answered
Boris left Anthropic for Cursor due to his admiration for Cursor's product and team, but quickly realized he missed Anthropic's mission-driven focus on AI safety, which he personally needs to be happy in his work.
AI is transforming software development at an accelerating, exponential pace; Claude Code now authors 4% of all public GitHub commits, and some advanced engineers, including Boris, have 100% of their code written by AI.
Boris believes that in a year or two, the need to understand the underlying code will diminish significantly, as AI will largely solve coding, making it less crucial for individuals to learn to code in the traditional sense.
The next big shift is AI agents starting to come up with ideas for what to build by analyzing feedback and bug reports, and branching out beyond coding to handle general computer-based tasks like project management and administrative duties.
AI is expected to transform roles adjacent to engineering, such as product managers, designers, and data scientists, and eventually any job that involves using computer tools, as agentic AI becomes more capable of acting on behalf of users.
Latent demand means observing how users 'abuse' or creatively adapt an existing product for unintended purposes, which reveals a strong underlying need. Building a dedicated product for this observed behavior often leads to significant success, as seen with Facebook Marketplace and Cowork.
Anthropic employs three layers of AI safety: mechanistic interpretability (studying model neurons), evals (laboratory testing in synthetic situations), and observing model behavior in the wild through early product releases to learn and improve alignment.
Boris plans to return to a lifestyle similar to his time in rural Japan, focusing on making miso, which teaches long-term thinking and patience, contrasting with the fast pace of engineering.
30 Actionable Insights
1. Build for Future AI Capabilities
Design and build AI products with the anticipated capabilities of models six months in the future, rather than current limitations. This forward-looking approach positions your product to “click” when more advanced models are released.
2. Bet on General AI Models
Adhere to “the bitter lesson” by consistently betting on more general AI models over highly specific or fine-tuned ones for most applications. General models typically outperform specific ones over the long term, despite initial performance gains from scaffolding.
3. Empower AI with Tools and Goals
Avoid overly strict workflows or excessive pre-context when building with LLMs; instead, provide the model with a set of tools and a clear goal, then allow it to determine the best course of action. This approach often yields better results than rigid orchestration.
4. Enable Unlimited AI Token Usage
For innovation, initially provide engineers with unlimited AI token access, rather than focusing on cost optimization. This freedom allows for experimentation with “crazy ideas” that might lead to breakthroughs.
5. Optimize AI Costs Later
Defer cost optimization for AI usage until an idea proves successful and begins to scale. Early optimization can stifle innovation, so prioritize proving the concept first, then find cheaper models or methods.
6. Automate with AI Agents
When faced with a task, consider if an AI agent can perform it, even if you could do it yourself. This principle encourages leveraging AI for automation to free up human time for higher-level work.
7. Adapt to Evolving AI Capabilities
Continuously update your mental model of AI capabilities, as models improve rapidly and significantly. Avoid getting stuck in old ways of thinking and actively explore what new versions can do.
8. Use Most Capable AI Model
Always opt for the most capable AI model available (e.g., Opus 4.6 with maximum effort) as it often proves more cost-effective and efficient. Its higher intelligence leads to faster task completion with fewer iterations and less human intervention.
9. Utilize AI Plan Mode
Initiate complex tasks using “plan mode” (e.g., by prompting the AI not to execute code immediately) to allow the model to outline its approach first. This ensures a well-thought-out strategy before execution, leading to more accurate one-shot completions.
10. Integrate AI for Multi-Tool Automation
Connect multiple AI tools and instruct the agent to perform complex, multi-step automations, such as syncing data between spreadsheets and messaging people on Slack. This enables comprehensive project management and reduces manual toil.
11. Run Multiple AI Agents in Parallel
Maximize productivity by running multiple AI agents or Co-work tasks concurrently. Initiate several independent tasks and allow the AI to work autonomously, freeing up your time for other activities.
12. Use AI for Idea Generation
Leverage AI tools like Claude to analyze feedback, bug reports, and telemetry data to proactively generate ideas for bug fixes and new features. This transforms AI into a co-worker that helps define what to build next.
13. Direct AI to Feedback Channels
Point your AI agent (e.g., Claude Code, Co-work) directly at internal feedback channels like Slack threads to automatically identify potential tasks and generate solutions. This streamlines the process of acting on user input.
14. Develop Generalist Skills
Cultivate a generalist mindset by crossing over multiple disciplines, such as combining engineering with product, design, or business acumen. This broader perspective helps in understanding complex problems beyond a single domain.
15. Encourage Cross-Functional Coding
Promote a culture where all team members, regardless of their primary role (e.g., PM, designer, finance), learn to code. This empowers individuals to unblock themselves and contribute more broadly.
16. Apply Common Sense & First Principles
Always apply common sense and think from first principles in your work, rather than blindly following processes or momentum. If something “smells weird,” question it, as this approach leads to better outcomes and avoids failures.
17. Prioritize Mission-Driven Work
Seek out work environments with a strong, resonating mission, such as safety in AI, as it can be crucial for personal happiness and job satisfaction. This deep alignment can be more important than building a cool product alone.
18. Understand Underlying Layers
To excel in your field, especially in AI, deeply understand the foundational layers beneath your immediate work (e.g., model mechanics for AI engineers, infrastructure for product engineers). This comprehensive understanding enables better work.
19. Under-resource New Projects
When starting new projects, intentionally under-resource them slightly to encourage creative problem-solving and force reliance on efficient tools like AI. This approach can lead to faster shipping and innovation.
20. Pursue Promising Threads
If you feel you’re “onto something” with an idea, even if its full utility isn’t immediately obvious, dedicate time and effort to explore it. This persistent exploration can lead to unexpected breakthroughs.
21. Maintain Human Code Review
Even with 100% AI-generated code and AI-powered reviews, retain a layer of human review for non-prototype code to ensure correctness and safety. This acts as a crucial checkpoint in the development process.
22. Rapid Feedback Loop
Respond to and address user feedback as quickly as possible, ideally within minutes, to make users feel heard and encourage more contributions. This rapid cycle fosters a culture of continuous improvement and engagement.
23. Prioritize Speed
Foster a culture of speed and urgency, encouraging teams to complete tasks today if possible. Utilizing AI tools can significantly accelerate this process.
24. Actively Experiment with AI
To succeed in the evolving AI landscape, actively experiment with new AI tools, embrace the bleeding edge, and overcome any apprehension. Dive in to understand their capabilities and limitations.
25. Design for Model’s Intent
When building AI products, observe what the model naturally “wants” to do or is capable of, and design products that facilitate these inherent capabilities. This “on distribution” approach leverages the model’s strengths.
26. Anticipate AI Tool Use & Longevity
When building for future AI, anticipate significant improvements in the model’s ability to use tools, interact with computers, and run autonomously for extended periods. Design products that leverage these expected advancements.
27. Auto-Accept AI Edits Post-Plan
Once an AI’s plan is reviewed and approved, confidently auto-accept its subsequent code edits. With advanced models like Opus 4.6, a good plan often leads to correct, one-shot execution.
28. Explore Diverse AI Interfaces
Experiment with various AI interfaces (e.g., desktop app, mobile app, Slack integration, web, IDE extensions) beyond just the terminal. Find the interface that best suits your workflow and preferences, as the core AI agent remains consistent.
29. Start Co-work with Simple Tool Use
For new Co-work users, begin by instructing the AI to perform simple tasks that involve tool usage, such as cleaning up your desktop, summarizing emails, or responding to top emails. This helps familiarize users with its agentic capabilities.
30. Engage Directly for User Feedback
Actively engage with users on platforms like Twitter to solicit feedback and identify bugs. Directly responding to user issues and questions can lead to rapid problem-solving and foster strong community engagement.
7 Key Quotes
A hundred percent of my code is written by Claude Code. I have not edited a single line by hand since November.
Boris Cherny
Coding is largely solved. I imagine a world where everyone is able to program. Anyone can just build software anytime.
Boris Cherny
The title of software engineer is going to start to go away. It's just going to be replaced by Builder and it's going to be painful for a lot of people.
Boris Cherny
I have never enjoyed coding as much as I do today because I don't have to deal with all the minutia.
Boris Cherny
The more general model will always outperform the more specific model.
Boris Cherny
Ask not what the model can do for you. Maybe, maybe it's something like this. Just think about how do you give the model the tools to do things? Don't try to over curate it. Don't try to put it into a box. Don't try to give it a bunch of context up front, give it a tool so that it can get the context it needs.
Boris Cherny
It just feels like this is 1% done and there's so much more to go.
Boris Cherny
3 Protocols
Tips for Building AI Products
Boris Cherny- Don't box the model in: Instead of strict workflows, give the model tools and a goal, then let it figure out how to achieve it.
- Bet on the more general model (The Bitter Lesson): Always favor more general models over specific, fine-tuned ones, as general models will likely outperform specific ones with future iterations.
- Build for the model six months out: Design your product for the capabilities of future models, even if it means the current product is 'uncomfortable' or not fully optimized today, to hit the ground running when new models are released.
Pro Tips for Using Claude Code Effectively
Boris Cherny- Use the most capable model: Currently Opus 4.6, with maximum effort enabled, as it often uses fewer tokens overall due to higher intelligence and efficiency.
- Use Plan Mode: Start almost all tasks in plan mode (Shift+Tab twice in terminal, or dedicated button in desktop/web app) to allow the model to plan with you before executing.
- Play around with different interfaces: Explore various form factors like the iOS/Android apps, desktop app, or Slack integration, as the same Claude agent runs everywhere, and find what feels right for your workflow.
Getting Started with Cowork
Boris Cherny- Start by having it use a single tool: Begin with simple tasks like cleaning up your desktop or summarizing emails.
- Connect multiple tools: Progress to tasks that involve connecting different tools, such as looking at emails and then sending Slack messages or updating a spreadsheet.
- Run multiple Cowork tasks in parallel: Kick off several tasks simultaneously and let them run unattended, allowing you to get a coffee while the agents work.