Skip to content

boris-creator-workflow.md

Latest commit

 

History

History
198 lines (131 loc) · 5.89 KB

boris-creator-workflow.md

File metadata and controls

198 lines (131 loc) · 5.89 KB
Raw

Boris Cherny's Production Workflow

Boris Cherny is the creator of Claude Code at Anthropic. This is how he actually uses it day-to-day — sourced directly from his public tips (Feb–Apr 2026).

The Core Philosophy

"Give Claude verification mechanisms. That's probably the most important thing. Testing improves final output quality by 2–3x."

Boris's entire workflow is built around one insight: Claude performs dramatically better when it can verify its own work. Not just running tests — systematic verification at every stage.

The 5+5 Parallel Setup

Boris runs 5 Claude sessions locally and 5–10 additional sessions on claude.ai/code simultaneously.

Local terminal sessions (5 worktrees):
├── Session 1: Feature A  (git worktree)
├── Session 2: Feature B  (git worktree)
├── Session 3: Bug fixes  (git worktree)
├── Session 4: Tests      (git worktree)
└── Session 5: Infra work (git worktree)

Cloud sessions (claude.ai/code, 5-10):
├── Long-running research tasks
├── PR reviews
├── Documentation generation
└── Background refactors

Why this works: Each session has a fresh context window, a clean context, and a specific mandate. No session is contaminated by another's work.

Model Selection (Boris's exact choice)

Opus 4.7 with extended thinking, for coding.

Counterintuitive: Opus is "larger and slower initially" but Boris found it:

  • Requires significantly less steering (fewer corrections)
  • Excels at tool use (bash, file editing)
  • Completes tasks in fewer turns
  • Ultimately faster end-to-end despite slower individual responses

The rule: Don't optimize for token speed. Optimize for task completion speed. A slower model that gets it right on turn 2 beats a fast model that needs 8 correction turns.

Plan Mode (Boris's workflow)

Boris starts every non-trivial task in Plan Mode.

1. Enter plan mode: Shift+Tab
2. Describe the goal + constraints
3. Let Claude reason through the approach
4. Review the plan — push back, refine
5. Approve the plan
6. Exit plan mode — Claude executes autonomously

The key: Once the plan is approved, let Claude execute with minimal interruption. The plan phase absorbs the steering cost. The execution phase should be largely autonomous.

Verification Loops (Boris's #1 tip)

This is the single most impactful technique Boris shares. See verification-loop.md for the full guide.

The short version:

# Instead of: implement → review manually
# Do: implement → auto-verify → fix → auto-verify → done

Configure hooks that run tests automatically after every edit. Claude sees the test results and self-corrects. Output quality improves 2–3x.

The Shared CLAUDE.md Pattern

Boris keeps CLAUDE.md in version control with the team.

The critical practice: After every session where Claude makes a mistake, add a rule to CLAUDE.md preventing that mistake in the future.

# Bad session: Claude keeps adding unnecessary console.logs
→ Add to CLAUDE.md: "NEVER add console.log debugging. Use the logger at src/lib/logger.ts"

# Bad session: Claude keeps creating separate files for utilities
→ Add to CLAUDE.md: "Utility functions go in the nearest existing utils.ts, not new files"

This turns every mistake into a permanent improvement. CLAUDE.md gets better with every session.

Slash Commands and Agents in .claude/

Boris stores all reusable workflows in .claude/commands/ and .claude/agents/, committed to git.

The principle: "If you've typed it more than once, it should be a command."

Examples from Boris's setup:

  • /verify — run the full test + lint + type check pipeline
  • /babysit-prs — automated PR review and merge checking
  • /batch-migrate — distribute a migration across worktrees
  • /voice — enable voice dictation mode

Voice-First Development

Boris primarily codes via speech rather than typing.

Why this matters for prompt quality:

  • Spoken prompts are naturally more detailed (easier to talk than type)
  • More context = better output
  • Faster iteration on complex prompts
# Enable voice mode
/voice

# Speak your prompt instead of typing it
# "Refactor the authentication middleware to use the new JWT library,
#  make sure to handle token expiry edge cases, and update the tests.
#  Follow the pattern in src/auth/session.ts."

Voice naturally forces you to include the WHY (you'd feel odd saying "refactor auth" out loud without explaining why). Better prompts, less back-and-forth.

Mobile Development

Boris reviews and makes code changes from his phone using the iOS/Android app.

Practical mobile workflow:

  • Review PRs on mobile during commute
  • Approve/reject plans while away from desk
  • Quick bug investigations
  • Check on running agent sessions

Not for heavy coding — for staying unblocked when away from your machine.

The /loop Pattern

For automated, repeating tasks:

# Check PRs every 5 minutes and handle routine tasks
/loop 5m /babysit-prs

# Run daily code quality check
/loop 1d /audit

# Keep watching a failing test
/loop 30s "run the failing test and try to fix it if it fails"

/loop + /babysit-prs is Boris's way of handling PR queue management without manual attention.

The --bare Performance Optimization

# Standard startup (normal speed)
claude

# Bare startup (up to 10x faster SDK startup)
claude --bare

Use --bare for scripted/automated Claude Code invocations where you don't need the full interactive UI. Significant speedup for CI pipelines and hook scripts.

Cross-Device Sessions

# Start a session on your desktop
# Get the session ID
/rename my-feature-session

# Continue from another device or terminal
/remote-control
# or
/teleport my-feature-session

Long-running tasks started on your desktop can be monitored or continued from any device.