← Back to library
PLAYBOOK High confidence

10 Tips from the Claude Code Team

Battle-tested workflows from Boris Cherny—Claude Code's creator—and his team. Parallel worktrees, evolved CLAUDE.md files, subagents, and the practices that ship 259 PRs in 30 days.

by Tacit Agent
ai-coding claude-code productivity workflows tooling
Evidence-Backed 2 sources · 2 high credibility

This analysis cites 2 sources with assessed credibility.

2 High
0 Medium
0 Low
View all sources ↓

TL;DR

Boris Cherny created Claude Code. In 30 days, he landed 259 PRs—497 commits, 40k lines added, 38k removed. Every line written by Claude Code + Opus 4.5. This thread documents the workflows that make that possible: parallel worktrees, evolving CLAUDE.md files, reusable skills, and subagents that keep context clean.


Quick Reference

# PARALLELISM
git worktree add ../feat feature-branch   # New worktree
cd ../feat && claude                       # Isolated session

# PLANNING
/plan                          # Enter plan mode
"Review this as a staff eng"   # Second opinion

# SUBAGENTS
"Spawn a subagent to write tests for this"
"Use a subagent to explore the auth module"

# PROMPTING PATTERNS
"Test this change before committing"
"Reimplement this more elegantly"
"Fix" + [paste entire Slack thread]

# CLAUDE.md ESSENTIALS
# - Add every mistake as a rule
# - Commit to git, share with team
# - Review weekly, prune outdated rules

Why This Source Matters

Boris Cherny isn’t a power user sharing tips. He built Claude Code as a side project in September 2024. When he describes workflows, he’s describing how the tool was designed to be used. When his team shares practices, they’re sharing what works at the source.

The 259 PRs stat isn’t marketing—it’s a public claim from someone whose reputation is tied to the tool’s success. The practices below are what enable that velocity.


The 10 Tips

1. Parallel Worktrees

“Spin up 3–5 git worktrees at once, each running its own Claude session in parallel.”

~/project/
main/
bug fixes
CC_
feature-auth/
auth system
CC
refactor-api/
API cleanup
CC
experiment/
spike
CC
[main] Working on bug fixes...

Setup:

# Create worktrees
git worktree add ../project-feature feature-branch
git worktree add ../project-refactor refactor-branch

# Each worktree gets its own Claude session
cd ../project-feature && claude

The insight: Claude Code is designed for parallelism. One session per worktree. Context stays clean. Ship multiple things simultaneously.


2. Plan First

“Begin complex tasks in planning mode before implementation.”

1Plan
Claude (Planner)
Creating plan...
>
2Review
Claude (Staff Eng)
>
3Build
Execute Plan
PLAN.md
1. Add auth middleware
2. Create session store
3. Implement login endpoint
4. Add token refresh

The insight: Planning mode isn’t a suggestion—it’s how the team actually works. Separate planning from execution.


3. CLAUDE.md Evolution

“Maintain a documentation file that evolves with corrections.”

MDCLAUDE.md0 rules
Empty file...

Example CLAUDE.md rules:

# Project Rules
- Never use `any` type in TypeScript
- Always run `pnpm test` before committing
- Use kebab-case for file names
- API responses must include `requestId` for tracing

The insight: The team shares a single CLAUDE.md for the Claude Code repo. Multiple updates per week. When Claude makes a mistake, it goes in the file. Error rates drop over time because the file accumulates institutional knowledge.


4. Reusable Skills

“Create custom commands committed to version control.”

ExamplesUse Case
Tech-debt commandsAutomated refactoring tasks
Slack/GDrive integrationsExternal system access
Analytics agentsData queries without context switching

The insight: Skills aren’t just shortcuts—they’re institutional automation. Commit them. Share them. Let the whole team benefit from one person’s prompt engineering.


5. Automated Debugging

“Paste bug threads directly to Claude with ‘fix’ instructions.”

PracticeWhy It Works
Slack integration enabledBug reports flow directly
Copy-paste entire threadsFull context, no summarization loss
Single instruction: “fix”Claude has everything it needs

The insight: Context switching is the enemy. Bring the bug report to Claude, not the other way around.


6. Advanced Prompting

“Challenge Claude to test changes, request elegant re-implementations.”

TechniqueResult
”Test this change”Verification before commit
”Reimplement elegantly”Forces simplification
Detailed specificationsReduces ambiguity, better output

The insight: Vague prompts get vague results. The team writes detailed specs—not because Claude can’t figure it out, but because specificity gets better output faster.


7. Terminal Setup

“Ghostty terminal with colored tabs, tmux for task organization, voice dictation.”

ToolPurpose
GhosttyFast, colored tabs for visual context
tmuxTask organization, persistent sessions
Voice dictationFaster than typing for long prompts

The insight: The environment matters. Visual separation between sessions prevents confusion. Voice input is faster for natural language.


8. Subagents

“Deploy subagents for focused task handling, keeping main context clean.”

Main Session
Architect
Clean context
Tests
Idle
Docs
Idle
Refactor
Idle
Spawn subagents for focused tasks. Main context stays clean.

When to spawn subagents:

  • Writing tests for code you just implemented
  • Generating documentation
  • Exploring unfamiliar parts of codebase
  • Any task that would pollute main context

The insight: Don’t pollute your main session with tangential work. Spawn a subagent. Let it handle the focused task. Return to clean context.


9. Data Analytics

“Leverage CLI tools like BigQuery directly within Claude Code.”

PracticeWhy It Works
BigQuery via CLIReal-time metrics
No SQL writingClaude handles the query
In-session analysisNo context switching to dashboards

The insight: If it has a CLI, Claude can use it. Bring data queries into your coding session instead of switching to separate tools.


10. Learning Mode

“Enable explanatory output, generate visual HTML presentations, create ASCII diagrams.”

PracticeWhy It Works
Explanatory outputUnderstand unfamiliar code
HTML presentationsVisual learning from code
ASCII diagramsArchitecture understanding

The insight: Claude isn’t just for writing code. Use it to understand code. Ask for explanations. Ask for visualizations. The investment in understanding pays off.


The 259 PR Context

These tips enable extreme throughput:

30 DAYSClaude Code + Opus 4.5
PRs
259
Commits
497
Lines +
40k
Lines -
38k

Every line written by Claude Code + Opus 4.5. The practices above are what make this sustainable.


What’s Transferable vs. What’s Unique

Transferable to Any Team

PracticeDifficulty
Parallel worktreesLow—just start using them
Plan mode firstLow—behavioral change
CLAUDE.md evolutionLow—start adding corrections
Reusable skillsMedium—requires initial investment
SubagentsMedium—requires understanding

Requires Infrastructure

PracticeRequirement
Slack integrationMCP setup, Slack access
BigQuery in sessionCLI tools, permissions
Voice dictationmacOS/accessibility setup

Requires Expertise

PracticeWhy
”Reimplement elegantly”You need to recognize elegance
Staff engineer review simulationYou need to know what a staff engineer catches
Detailed specificationsYou need domain expertise to specify correctly

The Tacit Angle

Several of these practices relate directly to session memory:

Auth
API
Tests
Bug
"Why did we add that middleware?"
TipMemory ChallengeHow Tacit Helps
CLAUDE.md evolutionKnowledge lives in file, not sessionsSession history shows why rules were added
Parallel worktreesContext scattered across sessionsUnified search across all sessions
Plan modePlans exist in session, then vanishPlans persist, become searchable
SubagentsSubagent context is lostSubagent sessions captured too

The 259 PR workflow generates massive session history. Without persistence, the why behind those PRs disappears. The code ships; the reasoning evaporates.


Recommendations

Start Here (Any Team)

  1. Use parallel worktrees — One session per worktree, 3-5 active
  2. Start CLAUDE.md — Begin with errors you’ve seen, evolve it
  3. Plan before implementing — Especially for multi-file changes
  4. Create one reusable skill — Start with your most common task

Level Up (Intermediate)

  1. Subagents for focused tasks — Keep main context clean
  2. Voice dictation for prompts — Faster for long instructions
  3. tmux for session management — Persistent, organized

Advanced (Infrastructure Required)

  1. Slack/external integrations — Bring bug reports to Claude
  2. CLI tools in session — BigQuery, analytics, whatever you use
  3. Staff engineer simulation — One Claude plans, another reviews

Confidence Assessment

ClaimConfidence
These practices work for the Claude Code teamHigh — first-party source
259 PRs in 30 days is realHigh — public claim, reputation stake
Practices transfer to other teamsMedium — depends on expertise level
Parallel worktrees are universally beneficialHigh — low cost, clear benefit
CLAUDE.md evolution reduces errorsHigh — mechanism is obvious

Sources & Provenance

Verifiable sources. Dates matter. Credibility assessed.

INDUSTRY High credibility
January 2026

10 Tips from the Claude Code Team ↗

Boris Cherny · X (Twitter)

"Battle-tested workflows from Claude Code's creator: parallel worktrees, evolving CLAUDE.md files, plan-first approach, reusable skills, subagents, and terminal setup optimizations."

INDUSTRY High credibility
January 2026

Claude Code Origin and Adoption ↗

Boris Cherny · X (Twitter)

"In 30 days: 259 PRs, 497 commits, 40k lines added, 38k removed. Every line written by Claude Code + Opus 4.5. Created as side project September 2024."