Skip to main content

Claude Code Remote

When you want full control an LLM with hands on your files.

Getting Claude Code working takes 30 minutes. Getting it working well takes continuous investment, but that investment compounds.

By the end you will have an AI that knows your standards, follows your voice, and applies your judgment automatically — without typing the same instructions every session.

Machine

*nix machines work best.

  • Mac
  • Ubuntu Dual Boot

See dev machines for more.

Accounts

AI

Infra

Install

  1. VS Code
  2. Nodejs

Open VS Code and create a project directory.

Start running a terminal in VS Code.

Install Claude Code:

curl -fsSL https://claude.ai/install.sh | bash

Run:

claude --version

Open Session

Navigate to your project folder in the terminal:

cd /path/to/your/project
claude

You will see the Claude Code prompt. Type anything — ask it to summarize the files in the folder, or just say "hello." This confirms it is working.

What you are looking at: an AI with access to your files. It can read, edit, create, and run commands. Nothing happens without you asking — but it can do a lot when you do.

Write CLAUDE.md

CLAUDE.md is a file at the root of your project. Claude reads it automatically at the start of every session. It is how you teach the AI how you work.

Create the file:

touch CLAUDE.md

Open it and add this minimum viable template:

# Project Instructions

## What This Is

[One sentence describing your project]

## How to Write

- [Your voice or style — e.g. "Short sentences. No jargon."]
- [Any formatting rules — e.g. "Use headers for sections. No bullet soup."]

## What to Avoid

- [Common mistakes you do not want repeated]
- [Any off-limits actions — e.g. "Never delete files without asking."]

## Context

[Two or three sentences about who this work is for and what success looks like]

Fill it in honestly. This file compounds — the clearer it is, the better every session gets.

Creative project (content, marketing, media): focus the "How to Write" section on voice, tone, and what good looks like. "Write like Hemingway. Concrete nouns. Active verbs. No filler." is a valid rule.

Code project: focus on patterns, what frameworks are in use, and what standards to enforce. The Config Architecture guide goes deeper.

Add Rule

Rules are always-loaded constraints — things the AI should never forget. They live in .claude/rules/ and are imported by CLAUDE.md.

Create the first rule:

mkdir -p .claude/rules

Create .claude/rules/voice.md:

# Voice

Write short. Use the active voice. No pleasantries ("Great question!").
No hedging ("It seems like..."). Direct answers only.

Add this line to CLAUDE.md:

@.claude/rules/voice.md

The @ prefix tells Claude to load that file every session. Now the voice rule is always active.

Test it: ask Claude to write a paragraph about the project. If it is still wordy or hedging, tighten the rule. Iteration is fast.

First Skill

A skill is a reusable workflow — a set of instructions Claude follows when invoked with a slash command. Skills turn repeatable work into one-line commands.

Claude Code ships with built-in skills. Try one:

/commit

This skill reads git changes and writes a commit message following best practices. No instructions needed — the skill contains them.

Other built-in skills to explore:

  • /fix-links — finds and repairs broken internal links
  • /review-pr — reviews a pull request and flags issues

Custom skills are also markdown files. Any instruction you type repeatedly is a candidate. See Skills for how.

Capability Map

Five built-in tool categories form the base of everything Claude Code can do.

CategoryWhat it does
FilesRead, edit, create, rename, reorganize
SearchFind by pattern, regex content search, codebase exploration
ExecutionShell commands, servers, tests, git
WebSearch, fetch docs, look up errors
Code intelligenceType errors, jump to definitions, find references (requires plugins)

Run Environments

EnvironmentWhere it runs
LocalYour machine — full file, tool, environment access
CloudAnthropic-managed VMs for offloaded tasks
Remote controlLocal execution controlled from browser or mobile

Extend Claude

Five extension types plug into different parts of the loop.

ExtensionWhatWhen it loads
SkillsReusable procedures with quality gatesOn demand — slash command or auto-match on description
MCPExternal data sources and APIsTool definitions deferred by default
HooksShell scripts on lifecycle eventsEvery matched event — deterministic, not advisory
CommandsSingle-prompt slash commandsRegistered at session start
Agent SDKCustom agents with full orchestration controlBuilt as separate processes

Multi-Agent Work

PatternHow
Sub-agentsFresh context windows — work doesn't bloat the parent, returns a summary
Agent teamsMultiple agents on different parts, lead agent coordinates and merges
Orchestrator-workersCentral agent delegates to specialized sub-agents via the Task tool

Context Primitives

PrimitiveWhat
CLAUDE.mdProject instructions loaded automatically every session
Auto memoryClaude saves learnings across sessions — first 200 lines loads per session
Auto-compactionSummarizes older messages when context fills; guided by "Compact Instructions"

Permission Modes

Toggle with Shift+Tab:

ModeBehavior
DefaultAsks before edits and shell commands
Auto-accept editsEdits files without asking; still asks for other commands
Plan modeRead-only tools only — creates a reviewable plan before execution
Auto modeBackground safety checks (research preview)

Every file edit is snapshotted before changes. Reverse with Esc twice.

Session Control

FeatureHow
Resumeclaude --continue — same session ID, pick up mid-task
Fork--fork-session — branch a new approach, preserve history
TeleportPull a web or mobile session into the terminal
WorktreesParallel sessions on separate branches — see Worktrees

Workflow Patterns

Five core agentic patterns, each built from different primitives.

PatternDescriptionBuilt with
Prompt chainingSequential phases with handoffs between stepsPlan mode + Skills
RoutingDifferent instructions for different contexts or stacksConditional CLAUDE.md
ParallelizationIndependent tasks run simultaneouslySub-agents
Orchestrator-workersCentral agent delegates to specialistsTask tool
Evaluator-optimizerOutput validated inside the same loopInline skills

Integration Points

SurfaceEntry
Terminalclaude CLI command
IDEVS Code and JetBrains plugins
ScheduledCloud tasks (Anthropic infra), desktop scheduler, /loop in-session
ChannelsTelegram, Discord, iMessage webhooks
CI/CDGitHub Actions, GitLab CI/CD
BrowserChrome extension

The Loop

Three phases blend continuously:

Gather context → Take action → Verify results

Context management — CLAUDE.md, memory, auto-compaction — determines how far the loop runs before it degrades. The harness (tools + context + execution environment) turns the model into an agent.

Cascade Starts

You now have:

  • A CLAUDE.md that orients every session
  • A rule that enforces your voice
  • Skills you can invoke for repeatable work

That is the beginning of the Cascade:

Mantra (your intention)
→ Rule (always enforced)
→ Hook (runs automatically on every action)
→ System (you never think about it again)

You are at Rule. The next step — hooks — are shell scripts that run when Claude takes certain actions. They are optional but powerful: a hook can automatically check that every file Claude edits passes a quality test.

The pattern compounds. The longer the loop runs, the more it learns your judgment. The more it learns your judgment, the less you explain.

What Next

To...Go to
Climb to Step 3 of the ladderHermes Onboarding
Understand the full config architectureConfig Architecture
Build your own skillsSkills
Use subagents for parallel workAgents
Choose the right AI modelLLM Comparison
See the Cascade concept in fullMantra

Context

Questions

What would have to be true about your CLAUDE.md for Claude to be genuinely useful after one session, without any additional instruction?

  • Which repeatable workflow would save the most time as a single slash command?
  • If the Cascade fully runs — Mantra → Rule → Hook → System — what does your work look like differently in six months?
  • Where does your judgment matter most today, and is that where the AI should learn first?