Most people barely scratch the surface of their $200/month Claude Max subscription. Here's the complete playbook for running at 100% usage — from basic settings to running 30 concurrent sessions overnight.
Most people are stuck at Level 1 or 2. The jump to Level 3/4 is where real value extraction happens.
Using claude.ai for questions, writing, and code snippets. One conversation at a time. Waiting for responses before asking the next thing.
Using Claude Code CLI in your project. Asking it to edit files, run tests, search code. Still one session at a time, waiting for each task.
Full tool permissions enabled. Large, complex prompts. Multiple parallel sessions using claude --dangerously-skip-permissions or proper allow-lists. Claude works while you review.
Using claude-fleet to run 10-30 concurrent Claude sessions. Tasks queued overnight. Remote access via OpenClaw. Reviewing outputs, not writing prompts.
The single biggest bottleneck for most people: Claude stops and asks for permission on every file read, edit, and command. Fix this first.
{
"permissions": {
"allow": [
"Bash",
"Read",
"Edit",
"Write",
"Glob",
"Grep",
"Web Fetch",
"Web Search"
]
}
}
Pre-approves all core tools so Claude can read files, edit code, run shell commands, search the web, and navigate your codebase without stopping to ask. This alone can 3-5x your throughput because Claude chains actions without waiting.
claude --dangerously-skip-permissions -p "your task prompt here"
You can also set permissions per-project in .claude/settings.json within your repo. This lets you allow npm test and npm run build without giving blanket Bash access everywhere. Add specific commands to the allow list: "Bash(npm test)", "Bash(npm run build)"
The more context and tools Claude can access, the bigger the tasks it can handle autonomously. These connectors turn Claude from a code assistant into a full-stack operator.
Search issues, create PRs, read repo contents, manage branches. Essential for autonomous PR workflows.
Read/write files across multiple repos. Critical for cross-repo audits and large-scale refactors.
Research APIs, documentation, competitors. Claude can read docs and apply them in context without you copy-pasting.
Post updates, read channels, notify on task completion. Useful for fleet status updates.
Query production data directly for analysis. Used in NammaYatri audit to analyze 50B+ ClickHouse rows.
Read tickets, create issues, link PRs to tasks. Makes Claude aware of project management context.
Run tests, build, deploy, git operations, package management. The most important tool for autonomy — Claude can verify its own work.
Check dashboards, query metrics. Used for observability analysis in platform audits.
{
"mcpServers": {
"github": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-github"],
"env": { "GITHUB_TOKEN": "ghp_..." }
},
"filesystem": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-filesystem", "/path/to/repos"]
},
"postgres": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-postgres", "postgresql://..."]
}
}
}
Each connector you add expands Claude's autonomy radius. Without GitHub MCP, Claude writes code but you manually create PRs. With it, Claude goes from code → commit → PR → done. Without database access, Claude guesses at data patterns. With it, Claude queries real data, finds actual bugs, and writes targeted fixes.
The secret weapon for hitting 100% usage. Run 10-30 concurrent Claude Code sessions with a single command.
A parallel task orchestrator for Claude Code. You write task descriptions in markdown files, and claude-fleet converts them into concurrent Claude Code sessions, each working independently on its own task.
Claude Max throttles per-session, not per-account (in practice). Running 10 sessions means 10x the token throughput. Running 30 sessions overnight means you wake up to 200+ completed tasks.
This is how Mags ran 403 tasks across 15 repos in 48 hours.
github.com/nammayatri/claude-fleetfleet-convert && fleet-run# 1. Write task files
cat > tasks/audit-backend.md << 'EOF'
Audit the backend codebase for:
- Security vulnerabilities (OWASP top 10)
- Performance bottlenecks (N+1 queries, missing indices)
- Error handling gaps (unhandled exceptions, missing retries)
- Code quality (dead code, TODOs, deprecated APIs)
Output a structured report with severity, location, and fix plan.
EOF
# 2. Create tasks for each repo
for repo in backend frontend mobile infra shared; do
cp tasks/audit-backend.md "tasks/audit-${repo}.md"
done
# 3. Convert and run
fleet-convert --tasks-dir tasks/ --output fleet-tasks/
fleet-run --concurrency 10 --tasks-dir fleet-tasks/
The final piece: a machine that runs 24/7 so you can queue tasks at night and review results in the morning.
10 PM: Queue 50-100 tasks via fleet-convert
10:05 PM: Start fleet-run with concurrency 20
Sleep.
7 AM: SSH in, check TUI dashboard
7:30 AM: Review outputs, raise PRs for the good ones
Want help setting up your own fleet? Mags is happy to remotely manage your Mac Mini or cloud VM to get claude-fleet running. SSH in, configure the environment, run a test batch, and hand it off to you.
Where most people leak usage capacity without realizing it.
If you're clicking "Yes" on every file read and edit, you're spending 60% of your time watching a spinner. Fix: Set up the permissions allow-list above. Pre-approve Bash, Read, Edit, Write, Glob, Grep.
Claude Max supports multiple concurrent sessions. Running one session is like having 30 CPUs and using one. Fix: Open 3-5 terminal tabs, each with its own Claude Code session on different tasks. Or use claude-fleet.
"Fix this typo", "add a semicolon", "rename this variable" — these burn a full message round-trip on work that takes 2 seconds manually. Fix: Batch small tasks into bigger prompts. "Audit this file for all typos, naming issues, and style violations. Fix them all."
Without a CLAUDE.md, Claude re-discovers your project structure, conventions, and preferences every session. That's wasted tokens on context-building. Fix: Write a CLAUDE.md in your repo root with architecture, commands, conventions, and key files. Claude reads it automatically.
Finishing one task, thinking about the next, typing it out. Every gap is wasted subscription time. Fix: Pre-write task descriptions in a file. When one finishes, paste the next immediately. Or use fleet for zero-gap execution.
Your subscription runs 24/7 but you only use it 8 hours. That's 67% waste. Fix: Set up a Mac Mini / cloud VM. Queue tasks before bed. Review in the morning.
Explaining your project, tech stack, and goals every time you start a new session. Fix: Use CLAUDE.md + auto-memory. Claude remembers across sessions. Also use /compact strategically to extend long conversations.
Simple projects with few files don't generate enough tool calls to burn through usage. Fix: Target large codebases (10K+ lines), multi-repo work, platform audits, or cross-cutting refactors. Claude shines on tasks with high complexity and many files.
Techniques that compound over time.
/compact to compress context mid-conversation--print flag for one-shot tasks (no conversation overhead)git diff to review, not reading Claude's explanationHow one person ran a full-platform audit of NammaYatri — India's largest open-source ride-hailing platform.
| Metric | Before Claude Max | With Claude Max + Fleet |
|---|---|---|
| PRs per day | 0.4 | 12+ (30x) |
| Repos covered | 1-2 | 15 |
| Reports generated | ~5/week | 268 total |
| Concurrent sessions | 1 | 10-30 |
| Hours active per day | 8-10 | 24 (overnight fleet) |
| Scope of work | Single feature/bugfix | Platform-wide audit + fixes |
It's not about typing faster or being a better programmer. It's about task decomposition + parallel execution + 24/7 runtime. Break the work into 400 independent tasks, let 30 Claude instances chew through them overnight, and review in the morning. The bottleneck shifts from "writing code" to "reviewing code" — and review is faster.
Start with Step 1: set up your permissions. Then try 3 concurrent sessions. Then try fleet.
claude-fleet Presentation → Get Setup Help →