Claude Environment

Interface reference for people new to Claude's ecosystem

June 2026 · Claude 4 + Fable 5
🌐 Web 🖥 Desktop ⌨️ IDE / CLI
Surfaces & Access Points
🌐
Web
claude.ai
  • Full chat UI in Chromium / Firefox
  • Projects with persistent context & instructions
  • Connectors (MCP apps) via OAuth
  • Google Drive file access integration
  • Artifacts panel — code, HTML, SVG, React previews
  • Deep Research & web search built-in
  • Styles & Skills configuration
  • Context window: 200K tokens (Free/Pro/Team) · 500K on Enterprise (some models)
🖥
Desktop
claude.com/download → macOS / Windows
  • Native app, full Web feature parity
  • Local file access without upload step
  • Screen capture & image attach via system menu
  • Desktop notifications & background sync
  • MCP servers via claude_desktop_config.json
  • Keyboard shortcut: ⌥ Space global hotkey
  • Cowork — desktop file & task automation
  • Context window: 200K tokens (Chat) · Cowork auto-manages via summarisation — no hard session cutoff
⌨️
IDE / CLI — Claude Code
code.claude.com · VS Code ext · JetBrains plugin
  • Agentic coding: reads, edits & runs code autonomously
  • Terminal-first: claude REPL or claude -p pipe mode
  • Full codebase context via CLAUDE.md
  • Slash commands, skills, hooks & MCP tools
  • Dynamic Workflows — coordinate hundreds of parallel subagents with adversarial verification
  • Auto mode safety — blocks destructive git/terraform commands unless explicitly requested
  • /fast toggle — switches to Fast mode (2.5× speed) mid-session
  • /cd — move session to a new working directory without restarting
  • VS Code: inline edits, sidebar chat panel
  • JetBrains: plugin for IntelliJ family IDEs
  • Chrome extension (beta) — browse & act on web pages
  • Computer use (preview) — cursor, click, type on-screen
  • Context window: 1M tokens for Opus/Sonnet on paid plans — Pro requires usage credits enabled; 200K for Haiku
Sidebar & Input Controls
Sidebar Navigation
✦  New chat
Workspace
Home
🗂Projects shared context
Starred conversations
Recents
💬Bookmark Themes v1 plan
💬Claude Code setup guide
💬Vulnerability scanning
Account
⚙️Settings
🎛Capabilities & Styles
👤Profile / Plan
Input Bar — "+" Attach Menu
📎Upload fileAll
PDFs, text, code, spreadsheets, Word docs — up to 32 MB per file
🖼Add imageAll
PNG, JPEG, GIF, WebP for visual analysis & multimodal tasks
🎨Use styleAll
Apply a saved writing style (tone, format, persona) to this chat
🧩Add skillAll
Attach a SKILL.md folder — expertise applied automatically
🔌Use connectorWeb
Connect an MCP app (Asana, GitHub, Slack, etc.) via OAuth
📁Google DriveWeb
Browse & attach Docs, Sheets, Slides directly from Drive
📸Screen captureDesktop
Snapshot any window or region — sent directly as an image
🔍Deep ResearchWeb
Multi-step web research — synthesizes dozens of sources
Model Selection
Available Models  Select via model picker (Web/Desktop) · --model flag or /model command (CLI)
FABLE 5
claude-fable-5 · most capable
Anthropic's most capable widely released model, for the most demanding reasoning and long-horizon agentic work. Thinking is always on.
Adaptive thinking1M context128K outputPremium cost
OPUS 4.8
claude-opus-4-8 · default
Most capable Opus. Complex reasoning, long-horizon tasks, and highest accuracy, for research synthesis, intricate architecture decisions, and multi-step agentic work.
Adaptive thinking1M context128K outputHighest costRecommended
OPUS 4.7
claude-opus-4-7 · previous Opus
Previous-generation Opus. Highly autonomous, with strong long-horizon agentic work, knowledge work, vision, and memory.
Adaptive thinking1M context128K output
OPUS 4.6
claude-opus-4-6 · older Opus
Earlier Opus generation with adaptive thinking, strong reasoning, and solid agentic performance.
Adaptive thinking1M context128K output
OPUS 4.5
claude-opus-4-5 · legacy, still callable
Legacy Opus, still available for teams pinned to this version for reproducibility.
200K contextLegacy
SONNET 4.6
claude-sonnet-4-6 · balanced everyday option
Smart and efficient, for everyday development, code generation, and analysis. Best balance of intelligence and speed for production workloads and interactive use.
Adaptive thinking1M context64K output
SONNET 4.5
claude-sonnet-4-5 · legacy, still callable
Legacy Sonnet, still available. A solid speed and intelligence balance for teams pinned to this version.
1M context64K outputLegacy
HAIKU 4.5
claude-haiku-4-5-20251001 · Fastest, lowest cost
Lightweight and fast, for high-volume tasks, real-time classification, and cost-sensitive pipelines.
Sub-second latency200K context64K outputHigh throughput
Markdown File Management — Project vs Environment Scope
📄 .md File Ecosystem Persistent context & instructions Claude reads every session — no prompt re-pasting needed ↗ code.claude.com/docs/en/memory
📁 Project Scope Lives in your repo · commit to share with teammates
CLAUDE.md ./CLAUDE.md (project root) git ✓
The primary memory file. Loaded at session start and re-injected after /compact. Put stack conventions, build commands, architecture decisions, and team standards here. Target under 200 lines — longer files reduce adherence.
CLAUDE.md ./src/api/CLAUDE.md (subdirectory) lazy load git ✓
Loaded only when Claude reads a file inside that subdirectory. Ideal for module-specific rules (e.g. auth layer security constraints, API versioning policy). Not re-injected after /compact until Claude revisits that folder.
CLAUDE.local.md ./CLAUDE.local.md (project root) gitignore
Loaded alongside root CLAUDE.md but not committed. Use for personal overrides, local API keys, machine-specific paths, or WIP notes you don't want in the shared repo.
.claude/rules/*.md .claude/rules/security.md, .claude/rules/api/*.md git ✓
Modular rules files — keeps project CLAUDE.md lean. Rules can be path-scoped with a path: frontmatter key so they only load when Claude works in matching directories. Ideal for monorepos.
.claude/skills/*/SKILL.md .claude/skills/pr-prep/SKILL.md git ✓
Project-scoped skills — reusable procedures that only load when invoked. Keep procedures here, not in CLAUDE.md. The skill folder can bundle templates, examples, and scripts alongside SKILL.md.
🌐 Environment / Personal Scope Lives in ~/.claude · applies across all projects
~/.claude/CLAUDE.md Global personal memory personal
Loaded in every project, every session. Store your universal preferences: preferred language, editor conventions, how you like responses formatted, and coding habits that span all your work.
MEMORY.md ~/.claude/memory/ (auto memory dir) auto-written
Claude writes this automatically as you work — capturing preferences, build commands, and patterns it notices. First 200 lines or 25 KB loaded each session. Browse and edit via /memory. Plain markdown; safe to prune.
~/.claude/skills/*/SKILL.md ~/.claude/skills/summarize-changes/SKILL.md personal
Personal skills available across all projects. Good for individual workflows: your preferred PR format, personal debugging checklist, or research approach. Not shared with teammates.
~/.claude/agents/*.md ~/.claude/agents/reviewer.md personal
Personal subagent definitions — specialised AI assistants with custom prompts and tool scopes. Available in all projects. Project-scoped equivalents live at .claude/agents/*.md and can be committed.
Enterprise managed Deployed via admin console (plist / HKLM / remote) org-wide
Org-level CLAUDE.md and settings pushed by admins. Takes precedence over all user and project settings. Appears in /status as "Enterprise managed settings". Individual users cannot override it.
Loading behaviour & tips
⚡ Load order
Enterprise → global (~/.claude/CLAUDE.md) → project root (CLAUDE.md) → project local (CLAUDE.local.md) → nested (lazy). All are in context before your first message.
🔁 After /compact
Project root CLAUDE.md is re-read from disk and re-injected. Nested subdirectory CLAUDE.md files are not — they reload next time Claude reads a file in that folder. Give critical rules in the root file.
📏 Size limits
Target under 200 lines per CLAUDE.md. Files beyond that consume context and reduce adherence. Offload procedures to Skills, use path-scoped rules, and ruthlessly prune anything Claude already does correctly by default.
📥 @-imports
@.claude/rules/api.md in a CLAUDE.md imports another file. Useful for organisation — but the imported file still loads at startup and counts toward context. Imports reduce complexity, not token cost.
What goes where — quick guide
CLAUDE.md (root)
Stack & language versions
Build & test commands
Coding conventions & style
Architecture decisions
Review checklist
Nested / Rules files
Module-specific constraints
API versioning rules
Security requirements per layer
Package-level conventions (monorepos)
Domain logic notes
Skills (SKILL.md)
Multi-step procedures
Repeatable workflows (PR prep, deploy)
Custom slash commands
Tasks with supporting templates
Things you keep re-pasting
~/.claude (personal)
Response format preferences
Universal coding habits
Auto memory (Claude-managed)
Cross-project personal skills
Local credentials & overrides
Claude Code — Slash Commands & CLI
⚡ Built-in Slash Commands Type / in session to browse 60+ commands
/init Generate CLAUDE.md project guide from codebase ↗ memory
$ /init → walks codebase, writes CLAUDE.md with stack, conventions & project structure
/compact Compress context with optional focus instructions ↗ commands
$ /compact focus on auth module → summarizes history, preserves auth context
/clear Clear conversation history, free up context window ↗ commands
$ /clear → blank slate; use before switching to an unrelated task
/cost Show token usage statistics for this session ↗ costs
$ /cost → Input: 42k tokens · Output: 8k tokens · Cache hit: 78%
/model Swap model mid-session without restarting ↗ model-config
$ /model claude-opus-4 → switches immediately, context preserved
/mcp List, connect & manage MCP server connections ↗ mcp
$ /mcp → shows connected servers; /mcp add github → OAuth flow
/security-review Analyze pending git diff for security vulnerabilities ↗ commands
$ /security-review → flags injection risks, auth gaps, data exposure in staged changes
/rewind Roll back conversation & code to a prior checkpoint ↗ checkpointing
$ /rewind → pick a message; code reverted, conversation branched from that point
/config Open Settings: theme, model, editor mode, output style ↗ claude-directory
$ /config → interactive UI; set Vim mode, dark theme, auto-compact threshold
/batch <task> Spawn parallel subagents across codebase (Skill) ↗ sub-agents
$ /batch migrate src/ from Solid to React → spawns one agent per module in parallel
/plan Enter plan mode — review before Claude executes ↗ commands
$ /plan fix the auth bug → Claude proposes steps; you approve before any edits
/review Deep read-only review of current diff — complements /code-review ↗ commands
$ /review → flags logic errors & style issues; no auto-apply. Use /code-review --fix to apply findings
🖥 CLI Shell Commands & Config Files
claude Launch interactive REPL session ↗ cli-reference
$ claude "explain this project" → opens session with initial prompt
claude -p "…" Non-interactive pipe mode — single prompt, then exit ↗ cli-reference
$ cat logs.txt | claude -p "summarise errors" → CI-safe, no REPL
claude -c Continue most recent conversation in current directory ↗ cli-reference
$ claude -c → picks up where you left off; no context loss on restart
claude -r "name" Resume session by name or ID with an initial prompt ↗ cli-reference
$ claude -r "auth-refactor" "finish the PR" → resumes named session
--model <id> Specify model at launch time ↗ model-config
$ claude --model claude-opus-4 → starts session with Opus 4
--print / --no-repl Print output then exit without entering REPL ↗ cli-reference
$ claude -p "list todos" --print → stdout-only for scripts & CI pipelines
CLAUDE.md Project instructions loaded automatically on every session ↗ memory
# CLAUDE.md → stack: Node 20 + TS, always run `pnpm test` before committing
.claude/commands/ Custom slash commands as Markdown files (now unified with Skills) ↗ skills
.claude/commands/deploy.md → invoked as /deploy during any session
.claude/settings.json Permissions allowlist, MCP config, tool rules ↗ claude-directory
{"allowedTools":["bash","read"],"mcpServers":{...}} → scoped to repo
hooks Shell scripts that fire on tool events (pre/post bash, edit, etc.) ↗ hooks
PostToolUse[bash] → run pnpm lint automatically after every shell command
env vars Control behaviour via environment variables ↗ env-vars
ANTHROPIC_MODEL=claude-opus-4 claude → override default model for a session
! shell cmd Run shell command directly inside a session without conversation overhead ↗ interactive-mode
! git status → runs directly; skips Claude's conversational parsing
🛠 Creating Your Own Slash Commands Custom commands are now unified with Skills — both approaches create /name commands ↗ code.claude.com/docs/en/skills
How to create a command
✦ Modern — SKILL.md (recommended) Legacy — .md file (still works)
1
Choose a scope — project (.claude/skills/) or personal (~/.claude/skills/). Personal skills apply to all your projects.
2
Create the skill folder — the directory name becomes the command name.
mkdir .claude/skills/pr-prep
3
Write SKILL.md — add YAML frontmatter with a description (used for auto-detection) then markdown instructions below the --- separator.
4
Test it — start a session and type /pr-prep, or describe the task and let Claude auto-invoke it when the context matches.
5
Commit to repo — project-scoped skills live in .claude/, so teammates get them automatically. Treat them as code.
SKILL.md anatomy
--- (YAML frontmatter) name: pr-prep description: Prepares a PR — runs tests, lints, writes summary. Use when user asks to open or prep PR. invocation: explicit --- ## Current diff !`git diff HEAD` ← dynamic ## Instructions 1. Run `pnpm test` 2. Run `pnpm lint` 3. Write a PR description covering: what changed, why, and any risks.
!`cmd` lines are dynamic context injection — Claude Code runs the shell command and inlines the output into the skill before Claude reads it.
Scope levels
🏢 Enterprise
managed settings (admin)
Pushed to all users in your org. Set via admin console.
👤 Personal
~/.claude/skills/<name>/SKILL.md
Available across all your projects on this machine.
📁 Project
.claude/skills/<name>/SKILL.md
Scoped to one repo. Commit it — teammates get it too.
📄 Legacy
.claude/commands/<name>.md
Still works. No frontmatter, no supporting files. Skill takes precedence if names collide.
Invocation & tips
Explicit — you invoke
/pr-prep
Set invocation: explicit in frontmatter. Claude won't auto-run it — you control timing.
Automatic — Claude invokes
"what changed in my diff?"
Omit invocation or set auto. Claude matches description to context and loads the skill on its own.
Arguments
/pr-prep fix auth bug
Text after the command name is passed as $ARGUMENTS — reference it inside SKILL.md instructions.
💡 Tips
Keep instructions procedure-based, not factual (facts belong in CLAUDE.md). Add supporting files (templates, examples) as siblings in the skill folder. Use /skills command to list all loaded skills.
Core Concepts — Developer Reference
Commands
IDECLI
Slash commands in Claude Code — 60+ built-ins like /init, /compact, /batch plus fully custom commands stored as Markdown in .claude/commands/. Custom commands are unified with Skills: both are invoked as /name; the Skill version takes precedence if both exist.
Example — custom command
# .claude/commands/pr-prep.md Run tests, lint, summarise changes for the PR description. → invoke as /pr-prep
🧩
Skills
WebIDEAPI
Reusable capability packages defined by a SKILL.md file plus supporting assets. Claude auto-detects and applies the skill when the context matches — no manual selection needed. Build once, runs on Claude.ai, Claude Code, and the API unchanged. Stack multiple skills for complex workflows.
Example — SKILL.md structure
skills/internal-comms/ SKILL.md # instructions examples/3p.md # format guides templates/ # reusable files
🔌
Connectors
WebDesktop
MCP-powered integrations with 100+ external services — Google Drive, Asana, GitHub, Slack, Salesforce, and more. Connect via OAuth in Settings → Connectors. Claude can read, write, and act through these tools inside any conversation or Project.
Example — using a connector
"Pull my open Asana tasks and group by project" → Claude calls Asana MCP, returns live task data
🛠
MCP
WebDesktopIDE
Model Context Protocol — open standard for connecting Claude to external tools, databases, and APIs. Two deployment modes: local stdio servers configured in claude_desktop_config.json, or hosted URL-based servers (SSE/HTTP). MCP servers expose tools, resources, and prompts that appear as commands via /mcp__server__prompt.
Example — Desktop config
// claude_desktop_config.json {"mcpServers": { "github": { "command": "npx", "args": ["@mcp/github"]}}}
🚀
Platform / API
API
Anthropic API — POST /v1/messages for building Claude-powered products. Now available natively on Claude Platform on AWS with IAM auth. Managed Agents support self-hosted sandboxes (public beta) and private MCP servers — your sandbox, Anthropic's orchestration. Also available via Amazon Bedrock and Google Vertex AI.
Example — Managed Agent sandbox
// agent loop on Anthropic // tool execution in YOUR infra providers: ["cloudflare", "daytona","modal","vercel"] → private MCP servers supported
🧠
Context Management
APIIDE
New context editing and memory tools that let Claude agents handle long-running tasks without hitting limits. Covers /compact for session compression, prompt caching to cut costs on repeated content, CLAUDE.md for persistent project context, and memory file patterns for agentic loops.
Example — prompt caching (API)
system: [{type:"text", text: "...10k token doc...", cache_control: { type:"ephemeral"} }] → 90% cost reduction on repeats
✍️
Prompt Engineering
WebIDEAPI
Anthropic's official prompting best practices. Key techniques: clear role framing in system prompts, XML tags for structured output, chain-of-thought via thinking blocks, explicit format instructions, positive + negative examples, and task decomposition for complex workflows.
Example — XML structured output
"Analyse this code. Reply using: <issues>...</issues> <suggestions>...</suggestions> <severity>low|med|high</severity>"
📂
Projects
WebDesktop
Containers for related conversations with a shared system prompt, uploaded knowledge files, and persistent context across sessions. The chat-interface equivalent of a repo-level CLAUDE.md. Team members on the same plan share a Project's context and history.
Example — system prompt in Project
"You are a senior engineer on the Payments team. Always follow our style guide (see attached docs) and prefer Stripe SDK over raw HTTP."