Getting Started with Claude Code & Cursor

Most engineers aren't ready for this move. The mental model has to shift first. You're no longer the one typing every line — you're the architect, reviewer, and orchestrator. The agent writes; you direct.

Before diving in, watch this. It's a game changer — a new way of thinking about development that makes the rest of this guide click.

Some resources mention Claude Code, but if you're using Cursor (or vice versa), the process is the same. The agentic loop is the same. Don't skip the foundational resources because you're on a different tool.

Things that help get better output:

• →Understanding rules, skills, and custom commands (Cursor) / agents and claude.md (Claude Code)
• →Understanding how to design an agentic loop

spec → architect → plan → implement → test → build → lint → verify

In agentic coding, if your plan isn't robust, the output won't be either. Front-load your effort on the plan. Here's a spec format that gives agents what they need to produce high-quality output:

### Goal
[What are you building and why?]

### Constraints
→ [Hard constraints — technology, time, compatibility]
→ [Use "→ NA" if none]

### What done looks like
→ [Concrete, observable outcomes]
→ [Testable acceptance criteria]

### What I'm NOT specifying
→ [Intentional gaps — let the agent decide]
→ [Use "→ NA – use this if there are no intentional gaps"]

The “What I'm NOT specifying” section is important — it signals intentional gaps where you trust the agent to decide. Use “→ NA” if there are none.

A good plan should have an HLD, LLD, and data flow — so you can see the bigger picture through architecture diagrams. Ask the agent clarifying questions until you're happy with the plan.

Then pick the first task and build it in a new agent session (Cursor) or continue in the same terminal window (Claude Code). Keeping context small matters.

Want to build features in parallel? Use git worktrees.

Cursor has built-in support for git worktrees — each agent safely operates on its own branch and working directory without conflicting with others.

Open multiple Claude Code / Cursor instances, each on a different worktree, and hop between features freely.

Even if you're moving to Claude Code — watch this. The thought process and the agentic workflow are the same across tools.

Still not convinced about the agentic process? Start with something measurable. An Xcode Build Optimization Skill by @twannl helped improve Xcode build times by 40–50%. The same approach applies to app launch times, scroll hitch rates, and other key metrics.

Understand how context windows work and how to keep them small for efficient token usage. This thread covers reducing Claude Code token usage by up to 60%:

claude.md, skills, rules, commands, and agents — what they are and how they fit together:

What should go in your claude.md and how to design an efficient agentic loop. We spent one week just setting this up to get it right — it's the most important bit:

By its creator — the authoritative guide on how to get the most out of Claude Code:

Claude code usage status line helper:

Notchy — converts your MacBook's notch into a Claude Code state observer:

# 1. Download the repo and go to the folder
# 2. Build the project
# 3. Locate Notchy.app inside the build folder
# 4. Remove quarantine attribute
sudo xattr -rd com.apple.quarantine Notchy.app

Follow the github links for spec driven development setup. The wizard scans your project and wires everything up.

That's a wrap

This covers everything I know — from the agentic shift to getting the most out of Claude Code and Cursor. I'll keep updating this as I find more resources. If this was useful, follow along for more.