Claude Code's Creator Reveals "Claude Cowork"'s Setup

January 23, 2026

42 min episode · 2 min read

Episode

42 min

Read time

2 min

AI-Generated Summary

Published Jan 24, 2026

Key Takeaways

✓Parallel Task Workflow: Run five to ten Claude sessions simultaneously across terminal, web, and mobile apps. Start tasks in separate tabs, let Claude develop plans, then switch between them to approve plans and enable auto-accept mode. This multiplies productivity by managing multiple AI agents rather than working sequentially on single tasks, fundamentally changing work from deep focus to orchestration.
✓Opus 4.5 Selection Strategy: Always use Opus 4.5 with thinking enabled despite being larger and slower than Sonnet. The smarter model uses fewer total tokens because it requires less steering and correction, making it faster and cheaper overall. Once the plan is solid, Opus 4.5 executes code nearly perfectly, eliminating the back-and-forth iterations required by smaller models.
✓Team Knowledge Base Maintenance: Maintain a shared Claude.md file checked into Git that the entire team updates multiple times weekly. Whenever Claude makes an error, immediately add correction instructions to prevent repeating the same issue. Use the GitHub action to tag Claude on pull requests to automatically update the knowledge base, implementing compound engineering principles where no correction happens twice.
✓Output Verification Setup: Always provide Claude with tools to verify its own work, like the Chrome extension for testing web applications or running test suites for code. Without verification capability, Claude operates like a blindfolded painter or engineer who never sees results. Self-verification dramatically improves output quality as the model iterates based on actual results rather than assumptions.
✓Plan Mode First Approach: Start most sessions in plan mode and iterate with Claude until the plan is comprehensive and accurate before writing any code. Switch to auto-accept edits mode only after plan approval, allowing Claude to execute without interruption. Good planning eliminates most errors upfront, making plan quality the primary determinant of code quality in the final implementation.

What It Covers

Boris, creator of Claude Code, demonstrates Claude Cowork's capabilities and best practices. He shows how the desktop app automates file management, spreadsheet creation, browser control, and email tasks through a virtual machine interface. Boris shares his personal Claude Code setup, including parallel task management, Opus 4.5 usage, and team knowledge base maintenance strategies.

Key Questions Answered

•Parallel Task Workflow: Run five to ten Claude sessions simultaneously across terminal, web, and mobile apps. Start tasks in separate tabs, let Claude develop plans, then switch between them to approve plans and enable auto-accept mode. This multiplies productivity by managing multiple AI agents rather than working sequentially on single tasks, fundamentally changing work from deep focus to orchestration.
•Opus 4.5 Selection Strategy: Always use Opus 4.5 with thinking enabled despite being larger and slower than Sonnet. The smarter model uses fewer total tokens because it requires less steering and correction, making it faster and cheaper overall. Once the plan is solid, Opus 4.5 executes code nearly perfectly, eliminating the back-and-forth iterations required by smaller models.
•Team Knowledge Base Maintenance: Maintain a shared Claude.md file checked into Git that the entire team updates multiple times weekly. Whenever Claude makes an error, immediately add correction instructions to prevent repeating the same issue. Use the GitHub action to tag Claude on pull requests to automatically update the knowledge base, implementing compound engineering principles where no correction happens twice.
•Output Verification Setup: Always provide Claude with tools to verify its own work, like the Chrome extension for testing web applications or running test suites for code. Without verification capability, Claude operates like a blindfolded painter or engineer who never sees results. Self-verification dramatically improves output quality as the model iterates based on actual results rather than assumptions.
•Plan Mode First Approach: Start most sessions in plan mode and iterate with Claude until the plan is comprehensive and accurate before writing any code. Switch to auto-accept edits mode only after plan approval, allowing Claude to execute without interruption. Good planning eliminates most errors upfront, making plan quality the primary determinant of code quality in the final implementation.

Notable Moment

Boris reveals that half his coding now happens on his phone using the Claude iOS app. He kicks off three to five Claude Code sessions each morning from bed while checking Twitter, then checks progress throughout the day. This mobile-first coding workflow represents a fundamental shift he never predicted a year ago when Claude Code launched in the terminal.

Know someone who'd find this useful?