Boris Cherny’s Claude Code Workflow: System-Level AI Coding Practices

Research Date: 2026-01-20
Video Source URL: https://www.youtube.com/watch?v=nWj1rdZ9wG8
Original Thread: https://x.com/boris_cherny/status/2007179832300581177 (January 2, 2026)

Reference URLs

YouTube Video: “You’re Using Claude Code WRONG!” - Universe of AI, January 11, 2026
X Post by @intheworldofai - January 11, 2026
Original Boris Cherny Thread - January 2, 2026
Thread Reader: Boris Cherny’s Claude Code Setup

Summary

Boris Cherny, the creator of Claude Code at Anthropic, published a detailed workflow guide on January 2, 2026 describing his day-to-day usage patterns. The workflow represents a paradigm shift from treating AI coding assistants as chatbots to treating them as integrated engineering systems. Universe of AI subsequently created a video breakdown emphasizing the key insight: effective Claude Code usage requires structural changes to how work is organized, not merely improved prompting techniques.

The workflow encompasses parallel session management, systematic error documentation, verification feedback loops, and tool integrations that enable Claude to participate in the full engineering process rather than operating in isolation.

Original Thread: Boris Cherny’s 13-Step Workflow

The following represents Boris Cherny’s complete workflow as published in his January 2, 2026 thread. His stated framing: the setup is “surprisingly vanilla” and there is no single “correct” way to use Claude Code.

1. Parallel Terminal Sessions

Boris operates five simultaneous Claude Code instances in terminal tabs, numbered 1-5. System notifications alert him when any instance requires input. This approach treats Claude instances as parallel workers rather than a single sequential assistant.

flowchart LR subgraph TerminalSessions["Terminal Sessions"] TabOne["Tab 1: Claude"] TabTwo["Tab 2: Claude"] TabThree["Tab 3: Claude"] TabFour["Tab 4: Claude"] TabFive["Tab 5: Claude"] end Notifications["System Notifications"] --> TabOne Notifications --> TabTwo Notifications --> TabThree Notifications --> TabFour Notifications --> TabFive TabOne --> WorkStreams["Parallel Work Streams"] TabTwo --> WorkStreams TabThree --> WorkStreams TabFour --> WorkStreams TabFive --> WorkStreams

2. Web and Mobile Sessions

Beyond terminal instances, Boris maintains 5-10 additional Claude sessions via claude.ai/code in the browser. Sessions can be handed off between terminal and web using background processes (&). He also initiates sessions from the Claude iOS app, particularly in the morning, checking progress later in the day.

This multi-device approach means Claude is continuously running rather than being opened and closed for discrete tasks.

3. Model Selection: Opus 4.5 with Extended Thinking

Boris uses Opus 4.5 with thinking mode enabled for all tasks, despite its larger size and slower response time compared to Sonnet. The rationale: Opus 4.5 requires less human steering and course-correction, resulting in net time savings.

4. Shared CLAUDE.md for Collective Learning

A repository-level CLAUDE.md file serves as collective memory. When Claude makes mistakes or produces suboptimal output, corrections are documented in this file for future reference. This creates persistent learning that benefits all team members and all future Claude sessions.

5. Automated PR-Based Learning

During code reviews, team members can tag @.claude in pull request comments to flag issues. A GitHub Action automates adding these learnings to CLAUDE.md, creating a continuous improvement loop.

flowchart TD PullRequest["Pull Request"] --> CodeReview["Code Review"] CodeReview --> TagClaude["Tag @.claude with Issue"] TagClaude --> GitHubAction["GitHub Action"] GitHubAction --> UpdateClaudeMd["Update CLAUDE.md"] UpdateClaudeMd --> FutureSessions["Future Claude Sessions"] FutureSessions --> ImprovedQuality["Improved Output Quality"]

6. Plan Mode First

Projects begin in Plan mode (accessed via Shift+Tab twice). Boris collaborates with Claude to refine the implementation plan before allowing file edits. Once the plan is satisfactory, he switches to auto-accept mode for execution.

This approach separates planning from implementation, reducing the risk of Claude pursuing suboptimal approaches.

7. Slash Commands for Common Workflows

Frequently used workflows (commit → push → PR creation, etc.) are codified as slash commands stored in .claude/commands/. These commands are version-controlled and shared across the team, reducing redundant prompting and ensuring consistency.

8. (Gap in Original Thread)

The original thread skips from step 7 to step 9. This gap has been noted by multiple observers.

9. PostToolUse Hook for Formatting

A post-tool-use hook automatically formats code after Claude generates it. While Claude typically produces well-formatted code, this hook prevents minor formatting issues from causing CI failures or requiring manual cleanup.

10. Permission Management Without Security Bypass

Rather than using --dangerously-skip-permissions (which bypasses all safety checks), Boris pre-approves commonly used safe commands via /permissions. These approvals are stored in .claude/settings.json and shared across the team.

This approach removes friction from safe operations while maintaining guardrails against dangerous commands.

11. MCP Tool Integration

Claude Code is configured to interact with real systems via Model Context Protocol (MCP) servers:

Slack: Post messages and notifications
BigQuery: Run analytics queries
Sentry: Retrieve error logs and monitoring data
Search: Access documentation and codebase search

These integrations enable Claude to access the same information human engineers use for debugging, review, and decision-making.

flowchart TD ClaudeCode["Claude Code"] --> ContextProtocol["Model Context Protocol"] ContextProtocol --> SlackIntegration["Slack Integration"] ContextProtocol --> BigQueryAnalytics["BigQuery Analytics"] ContextProtocol --> SentryLogs["Sentry Error Logs"] ContextProtocol --> DocSearch["Documentation Search"] SlackIntegration --> RealSystem["Real System Interaction"] BigQueryAnalytics --> RealSystem SentryLogs --> RealSystem DocSearch --> RealSystem

12. Long-Running Task Management

For tasks that extend over time, Boris employs several patterns:

Background verification agents that check Claude’s work
“Agent stop” hooks that pause execution at defined points
Plugins like ralph-wiggum for managing autonomy levels
Relaxed permissions in sandboxed environments for uninterrupted execution

13. Verification Feedback Loops (Most Critical)

Boris identifies verification as the single most important factor for output quality. Claude must have a way to verify its own work:

Running test suites
Checking UI behavior in browsers
Validating outputs against expectations
Using simulators or staging environments

He reports 2-3x improvement in output quality when Claude can check its own work. This improvement stems not from increased model intelligence but from tighter feedback loops.

Production Statistics (December 2025)

Boris published metrics demonstrating the workflow’s effectiveness:

Metric	Value
Pull Requests	259
Commits	497
Lines Added	~40,000
Lines Removed	~38,000
Time Period	30 days
Model Used	Opus 4.5
Human-Written Lines	0

Every line of code was produced using Claude Code with Opus 4.5.

Video Analysis: Key Themes

The Universe of AI video emphasizes several patterns that emerge from Boris’s workflow:

Chatbot vs. System Mental Model

The fundamental distinction: most users treat Claude Code as a chatbot (single-threaded, request-response, human-initiated), while Boris treats it as a system (parallel, continuous, self-verifying).

flowchart LR subgraph ChatbotModel["Chatbot Model"] HumanRequest["Human Request"] --> ClaudeResponse["Claude Response"] ClaudeResponse --> HumanReview["Human Review"] HumanReview --> NextRequest["Next Request"] end subgraph SystemModel["System Model"] ParallelClaudes["Multiple Parallel Claudes"] BackgroundExec["Background Execution"] SelfVerify["Self-Verification"] NotifyComplete["Notification on Completion"] ParallelClaudes --> BackgroundExec BackgroundExec --> SelfVerify SelfVerify --> NotifyComplete end

Friction Reduction Without Safety Compromise

Multiple workflow elements focus on removing unnecessary friction while preserving safety:

Pre-approved permissions for safe commands
Automatic formatting to eliminate manual cleanup
Slash commands for repetitive workflows
System notifications instead of active waiting

Collective Memory and Continuous Improvement

The CLAUDE.md file and GitHub Action integration create a learning system that improves over time. Mistakes are documented once and avoided thereafter, rather than being repeatedly corrected.

Verification as Quality Multiplier

The claimed 2-3x quality improvement from verification represents the most significant finding. This suggests that current AI coding quality limitations stem partly from missing feedback loops rather than model capability limitations.

Practical Implementation Considerations

Minimum Viable Workflow

For practitioners beginning to adopt this approach, the video suggests prioritizing:

CLAUDE.md documentation - Start recording mistakes immediately
Plan mode usage - Separate planning from execution
Verification setup - Ensure Claude can run tests or check its work

Scaling Considerations

The parallel session approach assumes:

Sufficient API rate limits or subscription tier
Independent work streams that do not require sequential coordination
Effective notification systems to manage attention across sessions

Team Adoption

Shared configuration files (.claude/settings.json, .claude/commands/, CLAUDE.md) enable team-wide consistency. New team members inherit accumulated learnings automatically.

Key Findings

Parallel Claude sessions (5+ concurrent) enable treating AI coding as a system rather than a chatbot
CLAUDE.md serves as persistent collective memory for team learning
Plan mode (Shift+Tab twice) separates planning from implementation, reducing wasted effort
Pre-approved permissions via /permissions remove friction without bypassing safety
MCP tool integration enables Claude to access real engineering systems (Slack, BigQuery, Sentry)
Verification feedback loops produce claimed 2-3x quality improvement
Opus 4.5 with thinking mode is preferred despite slower speed due to reduced steering requirements
December 2025 production metrics: 259 PRs, 497 commits, ~40K lines added in 30 days with 100% Claude-generated code

References

Boris Cherny’s Original Thread - January 2, 2026
Universe of AI Video - January 11, 2026
WorldofAI X Post - January 11, 2026
Thread Reader Summary - Accessed January 20, 2026