Who's Using
Claude Code
& How
A comprehensive survey of AI coding stacks from top practitioners — sourced from YouTube, podcasts, blogs, and GitHub configs.
Three Tiers of Usage
The consensus from 30+ sources: there are three distinct patterns for how top practitioners integrate AI into their coding workflow.
Dual-Tool
Cursor for daily flow and inline edits. Claude Code for hard problems — debugging, architecture, multi-file refactors. The best of both worlds.
Claude Code-First
CLAUDE.md as the configuration backbone. Custom skills, hooks, slash commands. MCP servers for the entire dev ecosystem. Multiple practitioners report not opening an IDE at all.
Multi-Agent Orchestration
tmux terminals, git worktrees, custom coordinators. Steve Yegge runs 3 concurrent Max accounts. The merge wall is the bottleneck — agents reserve files before editing.
Cursor is the best AI editor. Claude Code is the best AI engineer. Windsurf is the best value.
— Consensus from multiple comparison surveysWho Uses What
Click any card to expand. These are real configurations from practitioners who've shared their setups publicly.
Every line of committed code last month was AI-written. ~259 PRs, ~40K lines. ~5 releases per engineer per day. Hasn't opened an IDE in over a month.
Slash Commands
- /commit — runs linting, generates messages via Haiku (faster/cheaper)
- /pr — pull request workflow
- /feature-dev — spec → plan → todo → step-by-step execution
- /code-review — first-pass review on all PRs
- /security-review — runs on all PRs
Advanced Patterns
- Sub-agents for migrations: 10 parallel agents + 5 dedup agents
- Stop hook: "if tests don't pass, keep going"
- Notification hook: ping Slack when Claude needs permission
- Engineers write "diary entries" after each task
Key Insight
"Using plan mode can 2-3x success rates pretty easily if you align on the plan first."
"Claude Code is not a product as much as it's a Unix utility."
Built Gas Town — "Kubernetes for AI agents." A Mayor agent delegates to 20+ workers in tmux terminals. Runs 3 concurrent Claude Max accounts at ~$100/hr.
Architecture
- "Mayor" agent as foreman/interface for all tasks
- 20+ specialized worker agents in separate tmux terminals
- Agents push branches, open PRs, can autonomously merge
- Beads: SQLite for fast queries + JSONL for git-tracked source of truth
Key Insights
"Trust = predictability, not capability. Takes ~2,000 hours of daily use."
"Joel Spolsky's 'never rewrite' rule is now obsolete for a growing class of codebases."
Thesis
Current tools (Claude Code, Cursor) are "already obsolete." The future is agent orchestration dashboards — factory farming of code.
Uses every major AI coding tool. "Model musical chairs" — swaps freely when one gets stuck. Starts every feature with spec.md → plan.md.
5-Step Workflow
- Specs before code ("15-minute waterfall")
- Break work into small chunks with prompt plan files
- Extensive context via Context7 MCP or gitingest/repo2txt
- Choose the right model — swap freely when stuck
- AI across full SDLC: CLI agents, async agents, orchestration
Key Quotes
"Vibe coding is not an excuse for low-quality work."
"The human engineer remains the director."
Runs 2-3 async code research projects per day. Built his own CLI tools (files-to-prompt, llm). Prefers minimal MCP — custom shell scripts instead.
Workflow Phases
- Research: ask for options with usage examples, build prototypes
- Production ("Authoritarian" mode): exact function signatures, dictated choices
- Async research: dedicated GitHub repos with AGENTS.md, fire and forget
Philosophy
"If someone tells you that coding with LLMs is easy they are (probably unintentionally) misleading you."
"It's not about getting work done faster, it's about being able to ship projects that I wouldn't have been able to justify spending time on at all."
Costs
$0.11-$0.61 per session. A full feature build: $0.61, 17 minutes.
Created the Compound Engineering methodology and open-source plugin. Plan → Work → Assess → Compound. The most-referenced workflow system.
Custom Commands
- /workflows:plan — feature ideas → detailed implementation plans
- /workflows:work — execute with worktrees and task tracking
- /workflows:review — multi-agent code review
- /workflows:compound — capture learnings into CLAUDE.md
- /lfg — zero to production, fully autonomous
- /slfg — swarm mode for parallel execution
Plugin
Open-source: github.com/EveryInc/compound-engineering-plugin
Measured 3x increase in test output. 80%+ of code changes written by Claude Code. 20+ custom skills across PHP, TypeScript, React Native, Terraform.
Measured Productivity Gains
- Commits/week: 26-30 → 50
- Test LOC/week: 707 → 2,043 (3x)
- Total LOC/week: 3,975 → 5,947
- Prod deployments/week: 19-22 → 28
Configuration Architecture
- Root CLAUDE.md + backend/ + frontend/ + mobile/ CLAUDE.md
- Per-feature directories in git with implementation plans
- /clear between every subtask — never waits for auto-compaction
- Meta-skill: "implementing-skill" that generates other skills
"Don't create CLAUDE.md for every module — only where you observe comprehension failures."
The deepest public documentation of Claude Code usage. Custom status bar, system prompt patching (41% token reduction), voice input, multi-model containers.
Optimizations
- System prompt: 18K → 10K tokens (41% reduction)
- Lazy MCP via ENABLE_TOOL_SEARCH: true
- /clone, /half-clone for conversation forking
- "Cascade method": 3-4 tabs, sweep left-to-right
Multi-Model
Outer Claude Code coordinates Codex + Gemini CLI in separate Docker containers via tmux.
"Context length degrades performance — treat context like milk: fresh and condensed."
Runs --dangerously-skip-permissions every session. Switched from Cursor. Custom hooks auto-format with Prettier and TypeScript check on every edit.
Model Strategy
Opus until 50% cap, then manually switches to Sonnet "when Opus is having one of its moments."
Notable Achievement
Successfully edited an 18,000-line React component — says no other agent could handle it.
PR Reviews
Custom claude-code-review.yml focused only on bugs and security — stripped all verbose style comments.
Created the Context Refresh Cycle: /summ+commit → /clear → /freshstart. Five-file system: PRD.md, CLAUDE.md, PLANNING.md, TASKS.md, SCRATCHPAD.md.
Context Refresh Cycle
- /summ+commit — update TASKS.md and SCRATCHPAD.md, then commit
- /clear — reset context window
- /freshstart — read PLANNING.md, CLAUDE.md, TASKS.md, SCRATCHPAD.md
Cautionary Tale
GPU acceleration project failed after days of iteration — domain-specific hardware knowledge (CUDA's BUF algorithm) can't be substituted by AI iteration.
Most Popular MCP Servers
Ranked by frequency of mention across all 30+ sources. These are the extensions that keep appearing in every power user's config.
100% of code added in the last 30 days was written by Claude Code itself — approximately 259 PRs and 40,000 lines of code.
— Boris Cherny, on how the Claude Code team builds Claude CodeUniversal Workflow Principles
Seven patterns that appeared in every power user's setup, regardless of tools.
Plan Mode First
Shift+Tab before executing. Boris Cherny: "the most important thing you can do." Can 2-3x success rates.
CLAUDE.md as Configuration Backbone
Layered: global + project + directory. Start minimal. Add rules only when failures repeat. 50-100 lines per file max.
Fresh Context Per Task
/clear between tasks. New sessions. Handoff documents for continuity. "Treat context like milk: fresh and condensed."
Commit After Every Change
Git as save points. Worktrees for parallel work. Put "commit after every change" in CLAUDE.md for easy rewind.
Verify Everything
Treat AI output as unreviewed junior developer code. Playwright MCP for browser testing. Automated tests as safety net.
Sub-Agents for Parallelism
Dedicated agents for code review, security review, research, testing. Isolated context windows prevent implementation bias.
Match Model to Task
Opus for complex architecture. Sonnet for everyday coding. Haiku for boilerplate. Multiple users run 2-3 models simultaneously.
Orchestration Tools
For when one Claude Code instance isn't enough. These tools coordinate fleets of AI agents.
Gas Town
"Kubernetes for AI agents." Mayor + 20 workers in tmux. Beads for persistence. $100/hr. brew install gastown
Deep Dive →Multiclaude
"Brownian ratchet" — always pushing forward. Supervisor + subagents. Auto-merge if CI passes. Singleplayer or multiplayer mode.
Deep Dive →Compound Engineering
Plan → Work → Assess → Compound. Open-source plugin. Converts to Cursor/Kiro format. Most-referenced methodology.
Deep Dive →Claude Squad
Terminal app managing multiple Claude Code / Codex / Aider instances in separate workspaces.
Deep Dive →Conductor
Native Mac app. Visual dashboard with git worktree isolation, checkpoints, and rollback. Used by Addy Osmani.
Deep Dive →TSK
Rust CLI delegating tasks to agents in sandboxed Docker containers. Safety-first.
Deep Dive →Container Use
Docker-based dev environments for coding agents. Isolated workspaces for safe parallel execution.
Deep Dive →Cost Reality
What practitioners actually spend. The range is enormous — from $0.11/session to $100/hour.
| Setup | Monthly Cost | Who |
|---|---|---|
| Claude Code Pro (hits limits in ~1hr) | $20/mo | Casual users |
| Claude Code Max | $100/mo | Dzianis, Steve Sewell, most power users |
| Cursor Pro + Claude Code Max | ~$120/mo | Dual-tool setup (most common) |
| API direct (avg daily) | ~$180/mo | Boris's internal stats (~$6/day) |
| Claude Code Max x3 (multi-agent) | $300/mo | Steve Yegge |
| a16z estimate (10K context, 3 queries/hr) | ~$833/mo | Projected heavy enterprise usage |
| Gas Town at full speed | ~$100/hr | Steve Yegge (spikes) |
It's not about getting work done faster, it's about being able to ship projects that I wouldn't have been able to justify spending time on at all.
— Simon Willison