73 tips from 1600+ hours of usage

Master Claude Code in the Terminal

Every actionable technique to become a power user — from context engineering to parallel workflows and beyond.

~ claude
Why Terminal? Get Started
0
Tips
0
Categories
0
Hours of Usage
73 of 73
THE CASE FOR THE TERMINAL

Why Claude Code is vastly more powerful than Claude.ai

Even if you never write a single line of code, Claude Code in your terminal gives you superpowers that Claude.ai simply cannot match.

Claude.ai (Web)

  • Project files capped at ~50 files, size restricted
  • Memory limited to fixed small storage
  • No access to your local filesystem or tools
  • Cannot run scripts, commands, or automations
  • No persistent instructions across sessions
  • No subagents, no parallel processing
  • Cannot integrate with MCP servers or APIs
  • No skills, no slash commands, no hooks
  • Conversation is isolated — can't act on your world
VS

Claude Code (Terminal)

  • Unlimited folders, files — your entire filesystem
  • Unlimited memory via CLAUDE.md files
  • Full read/write access to your local machine
  • Run any script, any command, any automation
  • Persistent instructions via CLAUDE.md every session
  • Subagents for parallel research and execution
  • MCP servers for Stripe, GitHub, databases, more
  • Custom /skills, /slash-commands, hooks, headless
  • It acts, creates, modifies, and automates

Unlimited Folder Structure

Create any number of folders, subfolders, files. Organize projects, notes, research — no upload limits ever.

Infinite Memory via CLAUDE.md

CLAUDE.md files act as permanent memory loaded every session. Rules, preferences, context — across all levels.

Custom /skills & Commands

Build reusable workflows: /branch-commit-pr, /spec-developer, /mermaid-diagram. Share across projects.

MCP Server Ecosystem

Connect to Stripe, GitHub, databases, Chrome, Obsidian — any tool with an MCP server becomes a superpower.

Subagents & Parallelism

Spawn subagents for research, log analysis, code exploration — without polluting your main context.

Headless Mode + Cron

Run Claude Code in background with JSON schema output. Cron jobs for automated research and processing.

FROM WEB TO TERMINAL

How to get started

You don't need to be a developer. If you can open a terminal and type a command, you're ready.

1
Install Claude Code

Open your terminal and run: npm install -g @anthropic-ai/claude-code. If you don't have Node.js, install it from nodejs.org first.

2
Log in & launch

Type claude in your terminal and hit Enter. It opens a browser window to authenticate with your Anthropic account.

3
Navigate to any folder

Before launching, cd into any project folder. Claude Code now has full access to every file — no uploading, no limits.

4
Create your first CLAUDE.md

Create a CLAUDE.md file in your project folder. Write any instructions — this is your persistent memory loaded every session.

5
Use inside VS Code

Open VS Code, press Ctrl+` to open the terminal, type claude. Full access to your project with the editor's sidebar.

6
Start with simple tasks

Ask Claude to organize files, summarize PDFs, draft emails, create spreadsheets. No coding needed — just describe what you want.

01Mindset & Philosophy3 tips
01The only limitation is your imagination
Claude Code can do video editing, automate Wi-Fi reconnections, explore DNA data, automate dating apps, do taxes, and more. Think beyond coding.
02Continuously update your mental model
Don't assume old limitations still apply. What failed 2 months ago might work perfectly now. Retry things with every new model release.
03Keep a struggles document
Maintain a document of things Claude Code struggles with. When new models drop, retry those tasks — you'll be surprised how many now work.
02Context Window Management9 tips
04Don't front-load files
Let the model search the codebase itself instead of manually feeding it files. Front-loading fills up the context window and degrades performance.
05Modularize your codebase
Keep files under 500 lines. Ask Claude to identify files over 500 lines and modularize. Shorter files = less context consumed per read.
06Aggressively offload to subagents
Docs lookup, log analysis, and codebase exploration should go to subagents, not the main context. Add this rule to your user-level CLAUDE.md.
07Include a "why" when spawning subagents
Instead of "how auth works" say "how auth works for rate limiting". This filters signal from noise and reduces overlapping results.
08Stay in the 0-50% context range
Performance degrades significantly past 50%. Structure your workflow to stay in this optimal zone as long as possible.
09Audit context with /context
Regularly check what's loaded — memory files, skills, MCP servers. Delete unnecessary memory files and disable unused skills.
10Turn off auto-compaction
/config → autocompact → false. Compact manually for more control over what gets preserved and what gets dropped.
11Compaction safety checklist
Safe if: model is on-task, progressing clearly, compacted fewer than 3 times. If confused or already compacted too much — start fresh.
12Context anxiety is real
Models take shortcuts near context limits. Use up to 200K of a 1M window max. Add CLAUDE.md instructions to ignore context limits if needed.
03Claude Code Configuration6 tips
13Turn on verbose mode
/config → verbose → true. See the thinking process and token count. The reasoning traces reveal insights you'd never think of.
14Hierarchical CLAUDE.md files
Root user-level → project-level → personal local (not committed to git) → subfolder-level. Most important instructions go in the highest relevant one.
15Keep CLAUDE.md files short
They're progressively disclosed — only the most relevant ones load. Split by concern: backend rules, macOS rules, frontend rules in separate files.
16Use reference files
For custom vocabulary, package quirks, or domain knowledge — put them in separate files and reference from CLAUDE.md. Only loaded when relevant.
17Update CLAUDE.md on every mistake
When Claude makes an error, tell it: "find the relevant CLAUDE.md and add a rule to prevent this." Your instructions improve over time automatically.
18Prune CLAUDE.md with each model update
Newer models may not need old workarounds. Review and remove rules that are no longer necessary to keep your instructions lean.
04Terminal & Workflow Setup3 tips
19Use Ghostty terminal
The Claude Code team's terminal of choice. Set up shortcuts: Ctrl+T new tab, Ctrl+Shift+R rename, Ctrl+D split pane.
20Set up notification hooks
Permission dialog hook sends a system notification. Stop hook shows a bell icon on the tab. Never miss when Claude Code needs you across tabs.
21Configure the status line
/status-line — show model name, context battery (% full), git branch, unstaged changes. Essential situational awareness at a glance.
05Parallel Workflows4 tips
22Run 3-5 sessions per project
Like playing Factorio — there's always something to attend to. One agent has a question, another needs a plan review, another just finished.
23Run 10-12 sessions across projects
Spot bottlenecks across different projects. Improve your pipeline and slash commands based on patterns you see across sessions.
24Use voice dictation
Tools like Hyperwhisper are much faster than typing prompts. Dictate to all sessions, switch between them quickly. Massive productivity boost.
25Batch similar tasks together
Profile API fix in one chat, page redesign in another. Don't mix unrelated work in the same session — it causes context switching and worse results.
06Workflow Techniques3 tips
26Add multiple working directories
Copy implementations between projects. "Look at how Stripe is integrated in Project A, then copy into Project B." Avoids solving from first principles.
27Isolate features in a clean folder
When Claude is stuck in existing patterns, build the feature in an empty folder first. Once it works, integrate into the main project with fresh perspective.
28Enable sandbox + HTTP proxy
Whitelist allowed domains to prevent prompt injection from untrusted sites. Create a proxy script with a fixed set of trusted domains per task.
07Planning & Feature Development7 tips
29Always start in plan mode
Shift+Tab to enter plan mode. Even for small features. Describe what you want, let Claude explore the codebase, review the plan, then execute.
30Ask 20 clarifying questions
The default 4 questions isn't enough. Explicitly say: "Ask me 20 clarifying questions about how this should work." Better plans, fewer surprises.
31Use the spec developer skill
Turns a ~150-line plan into a comprehensive ~600-line plan covering every edge case. Gives you peace of mind before implementation begins.
32Use spec developer explore
Spins up 3-5 explore subagents in relevant parts of the codebase first, then interviews you with much better, context-aware questions.
33Plan in one session, execute in another
Use "yes, carry context and auto-accept edits" to transfer the plan. Preserves context budget for post-implementation adjustments.
34Save every plan to a tasks/ folder
Enables undoing features later even after 5 more features were added on top. Also useful for analyzing project evolution over time.
35Re-enter plan mode for follow-ups
If existing context is relevant to the next task, re-enter plan mode in the same session. If not, start a brand new chat.
08Prompting Techniques3 tips
36Speak the agent's language
Read the thinking trace to learn how Claude refers to things. Use "CDN download URL" not "download URL". Precise terminology = faster, better results.
37Use right-sized screenshots
Not too small (can't locate it), not too big (ambiguous what you mean). Include a zoomed-in crop alongside the full view for best results.
38Use competitor references
Show screenshots of competitor UIs as reference. A picture is worth a thousand words of prompting, especially for design tasks.
09Debugging Strategies6 tips
39Use /rewind liberally
Go back to a good point in the conversation instead of arguing. Bad decisions linger and pollute the context — rewind removes them cleanly.
40Document all attempts for tricky bugs
Save failed fixes to a file, start fresh, load the file. By attempt 4-5, Claude finds the root cause — and you can remove unnecessary changes.
41Add extensive logging
When stuck in a loop, add logs everywhere, reproduce the issue, then use a subagent to analyze the logs. Remove all logs after fixing.
42Upgrade log-analysis subagent model
If Haiku can't find the error in logs, upgrade to Sonnet. The more capable model catches subtle issues that cheaper models miss.
43Start fresh for persistent bugs
As context grows, agents favor repeating actions over synthesizing novel plans. A new chat often solves what 5 retries couldn't.
44Smallest possible fix (Occam's razor)
After fixing, remove all unnecessary changes from failed attempts. Keep only the minimal diff that addresses the root cause.
10Advanced Techniques17 tips
45Convert plans to task lists
Long plans become actionable task lists. Complete each in subagents, or spawn an entire agent team to parallelize the work.
46Use built-in worktrees
claude -w feature-1. Only when two parallel features might overlap — holds the codebase static for each during implementation.
47Skip worktrees for non-overlapping work
Backend API + homepage copy changes can happen on main branch simultaneously. Only use worktrees when files might actually conflict.
48Create custom skills for repeated prompts
Build reusable skills like /branch-commit-pr, /spec-developer, /mermaid-diagram. Share at user level.
49Use disable_model_invocation: true
Add to skill front matter for skills only you should trigger. Prevents Claude from auto-loading them and wasting context.
50Set model per skill
Use Haiku for simple step-by-step workflows, Sonnet for moderate complexity, Opus for deep reasoning. Save cost and latency.
51Use mermaid diagram generator
Understand data flows after feature implementation. Great for explaining to teammates how organization creation integrates with Stripe, etc.
52Headless mode for batch processing
Receipt validation, data research, cron-triggered skills — all in background. Claude Code works while you sleep, eat, or touch grass.
53Define JSON schema for headless output
Get structured, parseable data from headless runs. Feed into scripts, dashboards, or automations for a complete pipeline.
54$200/month plan = ~$2,000 API value
Your subscription gives roughly 10x the value in API credits. Process massive amounts of data through background headless instances.
55Verify implementation with explore subagents
After a big feature, tag the plan file and spin up 3-5 explore subagents to confirm everything listed was actually implemented.
56Cross-verify with external agents
Tag an external agent like Codex to independently confirm all changes from your plan were made. A second opinion catches what you might miss.
57Tick off plan items, compact, continue
For long plans: tell Claude to tick off everything done so far, compact the chat, then tag the plan file again. Keeps progress clear.
58Use a plan manager skill
Organize past plans with a dedicated skill. Track which features were planned, implemented, or reverted. Analyze project evolution.
59Use forked context for localization
Run localization syncing and repetitive tasks in a forked context window. Prevents polluting your main session with mechanical work.
60Maintain separate to-do lists
Keep one to-do list for yourself and one for Claude Code. Claude can handle scheduled checks like "review post performance in 3 days."
61Don't follow random CLAUDE.md hype
Ignore Twitter threads saying "add all this to your CLAUDE.md." Only add rules that solve problems you've personally encountered and verified.
11Useful Commands4 tips
62/copy — preserve formatting
Copy Claude's output with markdown formatting preserved. Paste directly into Slack, Notion, or docs with proper formatting intact.
63/export — save entire sessions
Export the full conversation to a file. Share with clients to show how you diagnosed an issue, or archive for future reference.
64/insights — self-improvement
Claude analyzes your last dozen sessions and suggests what you could be doing better. A personal coach for your Claude Code workflow.
65/rename & /resume
Name sessions for easy return. Also renames the terminal tab in Ghostty/Warp. Resume any previous session with full context preserved.
12Tools & Integrations8 tips
66Use well-known packages
TRPC, Next.js, etc. have more training data = better first-time results. For newer libraries, install official MCP servers or use Exa search.
67Claude + Chrome MCP extension
Form filling, web scraping anti-scrape sites, checking post performance, maintaining to-do lists, filling out tax forms — all via browser automation.
68Switch to Gemini CLI for visual tasks
Gemini models have better visual understanding for UI aesthetics. Use the right tool for the job — Claude for logic, Gemini for visual refinement.
69Use Obsidian with Claude Code
All your notes are markdown files Claude Code navigates natively. Generate backlinks, diagrams, HTML explainers. Obsidian CLI makes it more powerful.
70Resumable subagents for deep research
When torn between approaches, spin up subagents to explore each option. Resume any subagent later for more detail without starting over.
71Mix models strategically
Opus most of the time. Sonnet for quick questions. Haiku for well-defined workflows. Define the model at the top of each skill for automatic switching.
72Install official MCP servers for key libraries
For heavily-used libraries like Trigger.dev, install their official MCP server. Get Search Docs and debugging tools that compensate for limited training data.
73Use Exa MCP server for online search
Keep Exa MCP server enabled across all projects. It lets Claude search online for documentation of newer packages not in training data.
FAQ

Frequently Asked Questions

Common questions about Claude Code, answered.

What is Claude Code and how is it different from Claude.ai?
Claude Code is Anthropic's official CLI that runs directly in your terminal. Unlike Claude.ai, it can read your entire codebase, edit files, run commands, manage git, use MCP servers, spawn subagents, and maintain persistent memory via CLAUDE.md files.
How much does Claude Code cost?
Claude Code requires a Claude Pro subscription ($20/month) or Claude Max subscription ($100-200/month). The Max plan is recommended for power users — it provides roughly 10x the value in API credits.
What is context engineering in Claude Code?
Context engineering is carefully managing what's in Claude Code's context window. Keep files under 500 lines, offload to subagents, stay in the 0-50% range, and use CLAUDE.md for persistent context without consuming tokens.
What is CLAUDE.md and how should I use it?
CLAUDE.md is a markdown file read automatically for project context. It works hierarchically: global ~/.claude/CLAUDE.md, project-root, and folder-level. Keep them concise and update when Claude makes mistakes.
How many parallel sessions can I run?
You can run 3-5 parallel sessions per project and 10-12 across multiple projects simultaneously. Each operates independently with its own context window.
What are MCP servers?
MCP servers extend Claude Code by connecting to external tools: Chrome for browser automation, Exa for web search, and servers for GitHub, databases, filesystem access, and more.
How do I debug effectively with Claude Code?
Use /rewind to undo, document all failed attempts, add extensive logging, upgrade subagent models for complex bugs, start fresh when stuck, and always ask for the smallest possible fix.
What is headless mode?
Headless mode runs Claude Code non-interactively from scripts, cron jobs, or CI/CD pipelines. Pipe prompts in and capture structured JSON output for automated workflows.
All tips by Ray Amjad — Watch the full video