In this session, Meaghan Choi, Design Lead for Claude Code at Anthropic, demonstrates how to optimize development workflows using Claude AI through advanced features like work trees, custom prototyping skills, and automated pull request management. The video provides a deep dive into the internal practices used at Anthropic to speed up the shipping process while maintaining high design standards.
Key Takeaways
- Use Git work trees to run multiple Claude sessions in parallel without file system conflicts.
- Configure Claude Code with high context models like Opus for better understanding of complex codebases.
- Implement custom skills like /prototype to generate multiple UI options and preview them instantly in a browser.
- Utilize Loop mode and Auto mode to let the AI handle multi-step tasks without constant manual approval.
- Adopt the philosophy of offloading non-coding tasks such as CSS polish and PR chores to the AI.
Managing Concurrent Sessions with Work Trees
One of the most practical tips shared is the use of work trees. When developers work with AI, they often want to run multiple tasks simultaneously. However, running several instances of Claude on a single local repository can lead to file conflicts where one process overwrites the work of another. By using the work tree command, developers create isolated copies of their repository. This allows them to have several Claude sessions open in different windows, each working on a separate branch or task, effectively multiplying their productivity without technical friction.
Advanced Configuration: Context and Speed
To get the most out of Claude Code, Choi recommends specific configurations based on the complexity of the task. For large codebases, she suggests using the Opus model with a high context window, which allows the AI to understand the relationship between many different files at once. Coupled with Fast mode, this reduces the latency of responses, making the development process feel more like a conversation and less like a series of disjointed commands. This setup is particularly effective for large scale refactoring or adding complex features that span across the entire application architecture.
Custom Skills and the Prototyping Workflow
A standout feature of the demonstration is the /prototype skill. Rather than writing long prompts every time a UI element is needed, Choi uses a custom skill that tells Claude to generate several different implementation options for a feature. The AI then compiles these into an HTML file and opens them in a browser preview. This enables a designer or developer to visually inspect different versions of a component: such as different layouts for a dropdown or search bar: and then choose the best one for further refinement. This workflow moves the developer away from writing boilerplate code and toward higher level decision making.
AI as a PR Shepherd and Quality Monitor
Beyond writing code, Claude can be used as a shepherd for pull requests. Anthropic engineers use automated routines to have Claude review open PRs, address minor code review comments, and check for linting errors. This automation extends to quality monitoring. Choi demonstrates a routine that scans for front end changes in the repository. If it detects that a developer has merged UI changes without a designer being involved in the Slack or Google Doc discussions, it automatically generates a design review PR and notifies the developer. This ensures that even in high speed environments, the product remains polished and brand compliant.
Practical Applications
To apply these lessons, developers should start by exploring the CLI capabilities of Claude Code. Setting up work trees for different feature branches is a quick win for productivity. Teams can also begin building their own library of custom skills to automate repetitive parts of their specific tech stack. Finally, setting up an auto mode shepherd for the PR process can significantly reduce the amount of manual churn engineers face during the final stages of the shipping cycle. By treating the AI as a junior partner that handles polish and chores, senior developers can focus on architecture and user experience.
Frequently Asked Questions
What is the advantage of using Git work trees with Claude Code?
Git work trees allow you to check out multiple branches of a repository simultaneously into separate directories. When using AI tools like Claude, this prevents different AI processes from conflicting with each other on the same files, enabling you to work on three or four different features in parallel on one machine.
How does Loop mode function in this workflow?
Loop mode allows Claude to continue working through a task until it reaches a defined goal without stopping for user input at every step. This is useful for complex tasks that require multiple iterations of coding, testing, and debugging, as the AI can self correct and proceed until the objective is completed.
Can Claude handle complex product design decisions autonomously?
According to the demonstration, Claude and other large language models are not yet fully capable of high level design on their own. They are excellent at executing specified styles or generating options, but a human designer must remain in the loop to make the final calls on aesthetics, usability, and product direction.
What is the polish tool workflow mentioned in the video?
The polish tool workflow involves using the web version of Claude to send small, incremental CSS or UI fixes. This is often faster than spinning up a full development session for minor tweaks. These small fixes can be batched together into a single pull request by the AI to keep the repository history clean.