Claude Code: From New User to Power User

WorkflowSettingsEcosystemArchitecture

1. The Mental Model Shift

Stop thinking of Claude Code as a chatbot in your terminal. It's a general work operating system that happens to be terminal-native. According to Architecture, the creator built it from a belief in "terminal simplicity as the right AI interface" — the same bet TypeScript made against JavaScript tooling. A CPO at a $2.6B company uses it to plan his entire workday with a single command. A CEO reports it doubles productivity by unifying 6+ inboxes and running overnight todo lists (from Workflow). The shift you need to make is from "I'll ask it questions" to "I'll delegate entire workstreams."

The key metric isn't lines of code written. It's productivity per engineer — and eventually per person (from Architecture).

2. Your First 2 Hours: Context Architecture

The single highest-leverage thing a new user can do is invest time in context setup before writing a single prompt. From Settings: system engineering beats prompt engineering — 2 hours of context architecture means subsequent prompts can be 10 words long and still produce high-quality output. Most users start from scratch every session; that's the gap between generic and executive-level output (from Architecture).

Minimum viable setup (from Settings):

Critical CLAUDE.md rules:

Advanced setup: A three-tier _MANIFEST.md in each project folder (Tier 1: source-of-truth docs, Tier 2: domain folders, Tier 3: archive) gives Claude a prioritized loading order — it reads the most important context first and asks clarifying questions before starting (from Settings).

3. The Core Workflow: Research → Plan → Build

Every non-trivial task follows this loop (from Workflow):

  1. Research first: Run /last30days [topic] or equivalent research commands to ground the plan in current data. Claude can cross-reference meeting transcripts against prior strategy docs — compounding context.
  2. Plan always: Unless it's literally a one-line change, there's always a plan.md first. This is the rule that separates power users from everyone else. Plan files are the checkpoint that survives everything — context gets lost? Point a new session at plan.md and pick up exactly where you left off. 70 plan files and 263 commits is one power user's setup.
  3. Build from the plan: Execute. Run 4-6 parallel sessions — while one plans, another builds, another researches.
  4. Handover: When the context window fills, run /handover — Claude generates a summary of decisions, pitfalls, and lessons for the next session. This preserves institutional knowledge.

Plans accept images: Paste screenshots of bugs, error messages, mockups, or Slack threads directly with Ctrl+V and run plan. Claude writes a plan from the image (from Workflow).

Plans work for everything, not just code: strategy docs, competitive analyses, content calendars, campaign briefs, articles — same workflow (from Workflow).

4. Environment Setup (The Configuration That Changes Everything)

These settings unlock the power-user tier (from Settings):

Bypass permissions — the single most important setting:

{
  "defaultMode": "bypassPermissions",
  "skipDangerousModePermissionPrompt": true
}

Add to ~/.claude/settings.json. Without skipDangerousModePermissionPrompt, Claude asks for confirmation every session. Shift+Tab also toggles bypass mode.

Audio notifications — essential for parallel sessions:

{
  "hooks": {
    "Stop": [{"hooks": [{"type": "command", "command": "afplay /System/Library/Sounds/Blow.aiff"}]}]
  }
}

Know which session just finished without watching the terminal. Power users use Warcraft/Starcraft/Mario sounds (2.2K likes — the appetite for this is real).

Zed autosave at 500ms for a Google Docs-like collaboration experience:

{"autosave": {"after_delay": {"milliseconds": 500}}}

Claude's edits appear in Zed instantly; your typing is visible to Claude within a second. Ghostty + Zed split-screen is the recognized power-user terminal setup.

5. The Six Extension Mechanisms

Claude Code has six distinct ways to extend it (from Settings). Understanding the difference between them is what separates users from power users:

Mechanism What It Does When to Use
Skills On-demand knowledge: a SKILL.md that becomes a slash command Reusable workflows you run repeatedly
MCPs External tool connections (Slack, Granola, Linear, GitHub) Connect Claude to live data sources
Plugins Packaged workflows (shareable, installable via npx) Pre-built capability packs
Commands Single markdown files → slash commands (simpler than skills) Quick prompt shortcuts
Subagents Delegated focused tasks with restricted tools + cheaper model Parallel work, protect main context
Hooks Shell commands triggered on pre/post actions Audio notifications, logging, automation

Skills are the highest-leverage investment. They're portable across Claude Code, desktop app, and Cowork — investments aren't locked to one product (from Ecosystem). Anthropic open-sourced 11 domain-specific plugins (sales, finance, legal, data, marketing, support) as ready-to-fork starting points.

Structure skills with progressive disclosure: A monolithic instruction file restructured into a slim main file (table of contents) + separate reference files loaded on demand achieves 89% context reduction with no loss of functionality (from Settings).

6. Parallel Sessions and Async Work

Running 4-6 parallel Ghostty sessions is the workflow multiplier (from Workflow). Requires bypass permissions — sessions run autonomously:

Overnight autonomous work is the next step: plan a task, kick it off, sleep, review in the morning. The "multiplayer todo list that works overnight" pattern — AI processes the queue while you sleep, you review and reprioritize in the morning (from Workflow).

Scheduled tasks (Pro/Max plans): Run /web-setup to grant scheduled tasks access to repos. Tasks accept a repo, cron schedule, and prompt — fully async cloud execution, local machine doesn't need to be running (from Workflow).

Taskmaster solves the reliability problem for long-running sessions: it monitors and forces resumption when Claude prematurely stops, acting as a "meta-agent that supervises the agent" (from Workflow).

7. The /loop Command (No Coding Required)

For business users and non-technical professionals, /loop enables recurring agentic tasks with a cadence + task prompt (from Workflow):

8. Advanced Power-User Patterns

Specialized harnesses compound value as skills accumulate — designer, marketer, sales configurations tuned for specific professional workflows outperform one-size-fits-all agents (from Workflow). Garry Tan (YC president) uses Claude Code skills as his primary development workflow; gstack (6.4K likes) is his open-source skill pack at github.com/garrytan/gstack.

Inter-instance messaging: Multiple Claude Code instances can communicate — one session asks clarifying questions of another working on related code (from Workflow).

The DOM element selector: Claude Code Desktop supports clicking any element — sends tag, CSS classes, styles, surrounding HTML, screenshot, and for React apps: source file path, component name, and current props (from Settings).

Skills as state machines: Brad Feld's /start command is 1,400 lines with 15 steps. Treating it as a state machine with explicit input/output contracts per step revealed 12-22 seconds of removable overhead — the skill had accumulated "defensive machinery" from development that became redundant performance drag (from Settings).

Industry-standard grounding: Add "research the industry-standard approach to this problem, use it to guide yours" to prompts — forces Claude to ground solutions in established patterns, reducing novel-but-fragile implementations (from Workflow).

Research tool: Power users invest 1,200+ hours in Claude-based research workflows — AI papers, market analysis, competitive intelligence. Claude is becoming a primary research tool, not just a coding tool (from Ecosystem).

9. The Ecosystem to Know

From Ecosystem:

Sources cited: