The One Rule Claude Code's Creator Never Breaks

Boris works at Anthropic and created Claude Code, making him the go-to expert when people get stuck using the tool. He corrects a major misconception where users expect Claude to complete tasks 100% perfectly on the first try and get disappointed with minor errors. Boris emphasizes that humans work similarly and make mistakes too. The fundamental principle he advocates is always giving Claude a way to verify its work, such as through tests when writing Python code. This verification creates a feedback loop that instantly increases the quality of the final result.

Users can ask Claude which tests to implement, and it will provide tests for every feature it builds, even without users needing to understand the code details. For autonomous work, Boris recommends editing the Claude.md file, which gets sent into every Claude Code session, to instruct Claude to mention how it would verify work before starting any task. This ensures a constant feedback loop is maintained throughout the development process. The Claude.md file serves as persistent instructions that guide Claude's behavior across all sessions. Even non-technical users can benefit from this approach by having Claude suggest appropriate verification methods.

Boris explains that verification differs across domains and should be tailored to the specific type of work being done. For simple cases, verification might be as straightforward as running a bash command to check results. For UI-based verification, multiple approaches exist including Claude's own browser extension, which can navigate entire websites and verify everything automatically. For mobile app development, iOS and Android simulator MCPs are available for verification purposes. Beyond basic verification, Boris recommends verification-led techniques like test-driven development (TDD), where tests for new features are written first before implementation, and screenshots can be used to verify UI changes.

Boris emphasizes the critical importance of the Claude.md file and how it should be unique to each repository. This file should contain all basic information a newly created Claude session needs including the tech stack, basic project structure, code style and team conventions, and most importantly, things Claude shouldn't do. When errors occur in sessions, teams should document them in Claude.md to prevent repetition. Boris has made it a habit for his entire team to contribute to Claude.md multiple times per week whenever they observe Claude doing something incorrectly. Each team member has specific responsibility for maintaining particular sections of the file, ensuring comprehensive coverage and accountability.

For full-stack applications, Boris recommends a structured approach where each microservice has its own Claude.md file, with separate files for frontend and backend components. This modular approach prevents configuration bloat and keeps instructions relevant to specific contexts. However, Boris warns against bloating the Claude.md file with excessive information. His own Claude.md is approximately 2.5K tokens, which he considers an optimal size. He provided specific examples of what sections his team covers in their Claude.md file. The key is balancing comprehensiveness with conciseness to ensure Claude receives essential guidance without overwhelming context.

Claude Code offers different operational modes including plan mode for planning, auto-accept mode that automatically accepts file edits, and dangerously skip permissions mode where Claude has permission to run everything. Boris explicitly states he never uses the dangerous permissions mode, especially in production environments where a single terminal command could break the entire project. Instead, he configures Claude Code's permissions system to specify which commands are allowed automatically, which require confirmation before running, and which are completely denied. Users can add any rules they want and set them at the project level, and Claude will follow these guardrails consistently.

The .claude folder contains a settings.json file where all permission commands and configurations are documented, allowing Boris to share these settings with his entire team for consistency. He argues that Claude Code is one of Anthropic's most badly named products because it's not limited to coding tasks. Users have successfully generated full Anetian workflows, set up entire SEO pipelines, created different agents for analyzing business markets, and built various other non-coding systems. Claude Code functions as a comprehensive system for controlling many different workflows, not just software development. Boris describes it as an orchestrator that can coordinate multiple tools and processes simultaneously.

Boris explains that Claude essentially uses all of his other tools for him through integration and orchestration. He uses Slack with Claude Code via the MCP server, along with tools like BigQuery and Sentry, all accessed through their CLIs inside Claude Code. This demonstrates Claude Code's capability as a form of orchestrator that can coordinate multiple services and tools simultaneously. The speaker shares a personal example of using Claude Code for orchestration with Notion, where Claude connects via MCP to create databases, populate them, and understand the context of video ideas. This practical example shows how Claude Code can manage content workflows and handle complex database operations through natural language instructions.

Boris reveals his actual workflow for implementing features or tasks, which always begins by switching to plan mode and clearly describing the desired outcome. Once Claude generates a plan, he continues from there after thorough review. He argues that having a good plan is extremely important and that all back-and-forth discussion with the agent should happen during plan mode. Boris only starts the agent once he has verified that the plan is actually valid and comprehensive. After plan validation, he kicks off execution in auto-accept edits mode, not the dangerously skip permissions mode, maintaining safety guardrails even during automated work.

For working at scale, Boris runs five Claude Code sessions in parallel simultaneously to maximize productivity. While working with multiple sessions, he numbers his tabs to maintain organization and track which session is which. This numbering system is crucial because when he receives notifications from different tabs, he immediately knows which terminal tab they're coming from and can respond appropriately. This parallel workflow demonstrates how experienced users can leverage Claude Code's architecture to work on multiple independent tasks or features simultaneously. The numbered tab system is a simple but effective organizational technique for managing concurrent AI-assisted development work.

Boris heavily utilizes web sessions of Claude Code, where users connect the web version to their GitHub account, grant access to a demo repository, and Claude works on it in the cloud. His workflow involves using the background agents feature to put long-running tasks into the background, freeing him to work on other things. He even uses Claude Code on his phone, where Claude automatically performs all work in a cloud session, and when complete, creates a new branch and pushes all code changes for review. There's also a teleport command that brings cloud sessions back into the local terminal, allowing users to view and work on the code themselves. This truly exemplifies working with an AI agent as if it were a junior developer operating independently.

Boris reveals that he uses Opus 4.5 with thinking enabled for everything, despite it being a bigger and slower model that takes more time than Sonnet and significantly more time than Haiku models. He provides a compelling rationale: even though Opus 4.5 is slower per task, it has a much lower chance of making errors compared to smaller models. As a result, the overall time spent working with the model and steering it back on track is actually much lower with Opus 4.5. This counterintuitive insight suggests that the slower but more accurate model saves time in the aggregate by reducing the need for corrections and iterations. Quality and accuracy outweigh raw speed when considering the full development cycle.

Boris uses the Claude Code GitHub action, which is essentially a bot that can be called inside authorized GitHub repositories. He provides a strong use case for this integration during the pull request review process. Whenever he finds mistakes during PR reviews, he simply asks Claude to add those mistakes to the Claude.md file automatically. This creates a continuous improvement loop where common errors get documented and prevented in future development. For long-running tasks, Boris employs three different approaches to maintain productivity. The first approach involves asking Claude to verify its work using different tests that Claude Code itself creates, then putting those tasks into the background to run independently.

Boris's background tasks are not cloud agents but local background processes that run silently and report back to the main agent once completed. The second approach for long-running tasks uses a stop hook to automatically trigger verification when Claude stops outputting, ensuring no work completes without validation. The third option is based on the Ralph Wiggum plugin, which the video creator covered extensively in another video. The Ralph plugin removes humans from the loop entirely during UI development by automatically testing and verifying that the UI Claude Code produces is actually correct. This represents the ultimate automation in verification, where Claude can develop, test, and validate interface changes without human intervention at each step.

Continuing the verification pattern, Boris uses linters and formatters for various programming languages to verify the format of code being written by Claude. He acknowledges that while AI models do generate correct code, these formatters help clean up the remaining 10% of issues that might exist. This formatting verification is particularly important to prevent errors when code runs in CI (Continuous Integration) as part of their GitHub integration pipelines. Proper formatting ensures that code meets team standards and passes automated checks before merging. The combination of functional correctness from Claude and formatting perfection from automated tools creates production-ready code with minimal manual intervention.

Boris reveals that his own workflow is surprisingly vanilla despite his expertise, challenging the assumption that power users need complex setups. This applies to most users who don't need to overcomplicate everything in Claude Code to be truly productive. The simple built-in features already provide about 90% of the productivity users are trying to achieve. This insight is valuable for beginners who might feel overwhelmed by advanced configurations. Boris emphasizes that mastering the fundamentals and built-in features is more important than creating elaborate custom setups. The next sections will cover these simple but powerful features that have been in Claude Code for a long time but many beginners don't know well.

Boris discusses /commands (slash commands) which he uses for his repetitive workflows, which he calls his 'inner loop' workflows - tasks he repeats many times daily. He provides a simple example of a GitHub /command that he uses frequently for common GitHub operations. All of these /commands are stored inside the .claude folder in a command subfolder, which means they can be committed to git version control. This is exactly how Boris shares his custom commands with his entire team, ensuring everyone has access to the same productivity shortcuts. Custom commands reduce repetitive typing and standardize common operations across the team, making collaboration more efficient and consistent.

Another powerful built-in feature is subagents, which people have extensively utilized for various use cases like market researchers or product validators. However, Boris takes a more focused approach, using subagents to automate simple workflows that require completely different instructions from the main agent. Most of his subagents are used to verify that the architecture of his code is correct, to refactor code that's been written, and to validate that final builds actually work. This specialization allows different subagents to have specific expertise and instruction sets optimized for their particular verification or transformation task. By delegating specific responsibilities to subagents, Boris creates a modular system where each component excels at its designated function.

The video includes a sponsored segment for Automata, the company behind the channel. After teaching millions of people how to build with AI, they started implementing these workflows themselves and discovered they could build better products faster than ever. Automata helps bring ideas to life by building apps or websites for clients who have great ideas but lack a technical team. They position themselves as a technical co-pilot, applying the same workflows they've taught millions directly to client projects. The service aims to turn concepts into real, working solutions without the headaches of hiring or managing a development team. Interested viewers are invited to reach out at hello@automata.dev to accelerate their ideas into reality.

The video concludes with a brief closing segment where the creator thanks viewers for watching. He mentions that viewers can support the channel and help them continue making similar educational videos by using the Super Thanks button below the video. This monetization option allows viewers to directly contribute to the channel's sustainability. The creator ends with his standard sign-off, thanking the audience again and promising to see them in the next video. This closing maintains the channel's approachable and appreciative tone while encouraging community support for continued content creation.