How Boris Uses Claude Code
Boris Cherny -- the creator of Claude Code -- describes his personal workflow in a Twitter/X thread. His approach is "surprisingly vanilla," which is itself the point: the tool is designed so you can use it, customize it, and hack it however you like. The most important takeaway: verification is "probably the most important thing to get great results" and can "2-3x the quality of the final result."
Key Quotes
"We intentionally build it in a way that you can use it, customize it, and hack it however you like. Each person on the Claude Code team uses it very differently."
"Probably the most important thing to get great results: verification. It can 2-3x the quality of the final result."
Key Themes
Parallel Sessions at Scale
Cherny runs 5 Claude instances simultaneously in his terminal (numbered tabs 1-5) with system notifications, plus 5-10 additional instances on claude.ai/code and the iOS app. He teleports sessions between platforms and hands off work between local and web environments. This is a much higher degree of parallelism than most users attempt.
Model Choice
He exclusively uses Opus 4.5 with thinking for everything, reasoning that superior steering and tool-use capabilities make it faster overall than smaller models requiring more correction. Bigger model, fewer retries.
Institutional Knowledge via CLAUDE.md
Teams maintain CLAUDE.md files checked into git repositories, where the team collectively documents Claude's errors and adds preventive instructions. This is the same pattern described in CLAUDE.md (Universal) -- the file becomes a shared institutional memory that persists across sessions and team members.
Plan Then Execute
Sessions typically begin in Plan mode (shift+tab twice), enabling back-and-forth refinement before switching to auto-accept edits mode for implementation. This matches the pattern in Trycycle and How to Write a Good Spec for Agents -- invest in planning before turning the agent loose.
Slash Commands and Subagents
Custom commands stored in .claude/commands/ automate frequent workflows. Example: /commit-push-pr uses inline bash to pre-compute git status, reducing model back-and-forth. Specialized subagents like code-simplifier and verify-app automate post-generation tasks.
Permission Management
Rather than --dangerously-skip-permissions, he pre-allows safe bash commands via /permissions, with common configurations shared team-wide in .claude/settings.json. This is the sane middle ground between YOLO mode and constant permission prompts.
MCP Integration
Claude Code integrates with external systems through MCP servers: Slack posting and searching, BigQuery queries for analytics, Sentry error log retrieval, Chrome extension testing for UI verification. This is the tool-use pattern that makes coding agents more than just code generators.
Verification is the Force Multiplier
Cherny's verification approach: Chrome extension-based browser testing, automated UI interaction, test suite execution, domain-specific validation. For long-running operations, he uses verification agents, Stop hooks, or the ralph-wiggum plugin. The claim that verification 2-3x quality is consistent with Feedback Loop is All You Need -- the feedback loop, not the generation step, determines output quality.
Code Quality via Hooks
A PostToolUse hook handles code formatting after generation, addressing the remaining 10% of formatting concerns to prevent CI failures. This connects to Pre-Commit Lint Checks -- automated enforcement beats instructions.
Critical Analysis
The value here is provenance. When the inventor of a tool explains their workflow, you're getting design intent, not just one user's adaptation. Several patterns stand out as deliberate design choices rather than emergent workarounds:
- Plan mode exists because it works. The shift+tab workflow isn't accidental -- it reflects the observation that refinement before execution beats correction after.
- CLAUDE.md is team memory. The creator uses it as collective error documentation, which suggests that's its intended purpose, not just a prompt engineering trick.
- Verification over generation. The 2-3x quality claim positions verification as more important than the initial generation -- the agent's first draft is the starting point, not the deliverable.
- Parallelism is normal. Running 5-15 instances simultaneously suggests the tool was designed for this. Compare with Addy Osmani's Workflow (power user perspective) and Claude Code on the Go (mobile-first adaptation) for different scaling approaches.
The gap: Cherny doesn't discuss failure modes or situations where Claude Code struggles. The creator's perspective is inherently optimistic. For the complementary view -- where agent-driven development breaks down -- see The Mythical Agent-Month and Slowing the Fuck Down.
Sources: How Boris Uses Claude Code Last updated: 2026-05-14