Look anything up.
A plain-language manual: every AI term, every command you can type, what each Desktop action maps to, and everything Claude can do.
A
- A2AAgent-to-Agent protocol · Agent2AgentA2A (Agent-to-Agent) is an open standard that lets separate AI agents discover each other, talk, and split up work — even when they were built by different companies on different frameworks. If
MCPis how one agent reaches its tools, A2A is how two agents reach each other. Each agent publishes anAgent Carddescribing what it can do, and others hand it tasks through a shared, defined lifecycle.ai - AEOAnswer Engine Optimization · Generative Engine Optimization · GEOAEO (Answer Engine Optimization) is writing your content so AI answer engines — ChatGPT, Claude, Perplexity, Google's AI Overviews — quote it directly when someone asks a question. Where classic SEO competes to rank in a list of blue links, AEO competes to be the answer the AI reads out. In practice that means answer-first writing, clear structure, real FAQ-style Q&A, and machine-readable facts a model can lift verbatim.ai
- AgentAn AI that doesn't just answer — it takes actions in a loop to reach a goal: read a file, decide the next step, run a command, check the result, repeat. Claude Code is an agent. A plain chatbot only talks; an agent does things on your behalf.ai
- Agent SDKClaude Agent SDKA toolkit for TypeScript or Python that gives your own program the full Claude Code engine as a library — the agent loop, tools, permissions, context management, MCP, hooks and subagents — instead of a chat window around it.ai
- Agentic AIAgentic workflowsAgentic AI is an AI that works out the steps itself toward a goal — making a plan, taking actions with tools, checking the result, and adjusting — instead of doing the single thing you spelled out. A plain chat reply answers your one question and stops; an agentic system keeps going until the job is actually done, looping through act-and-check on its own. Claude Code is agentic: you give it a goal, and it reads files, runs commands, and fixes its own mistakes along the way.ai
- Agentic codingAI coding agentAgentic coding is using an AI that actually does the software work for you — reads your files, writes and edits code, runs commands, checks the result, and fixes its own mistakes — instead of one that just hands you a snippet to paste. You describe the outcome you want in plain language; the agent works out the steps and carries them out.
Claude Codeis an agentic coding tool, and you don't have to be a programmer to direct one.ai - AGIArtificial general intelligenceA hypothetical AI that could match a capable human across basically any task — learning, reasoning, and adapting broadly — rather than being strong at a narrow set of things. Today's models, impressive as they are, are not AGI: they're remarkably good in specific domains but still have real gaps and blind spots. AGI is a goal and a topic of debate, not a product you can buy.ai
- AIartificial intelligenceSoftware that does things we used to assume needed a human: understanding language, recognizing images, making decisions. It's a broad umbrella — today's most useful kind for everyday work is the
LLMbehind tools like Claude.ai - AlignmentAI safety · GuardrailsThe work of making an AI's behavior match what people actually intend and value — being helpful and honest, and declining requests to do real harm. Guardrails are the practical limits that enforce this in day-to-day use: the boundaries that keep a capable system from being misused or going off the rails.ai
- APIApplication Programming InterfaceHow one program asks another to do something — a defined set of requests a service agrees to answer. When an app shows you the weather, it's quietly asking a weather service through its API. It's a menu of what you're allowed to order, and the format for ordering.software
- Ask permissionspermission mode · Ask permissions mode
Ask permissionsis the mode inClaude Code Desktopwhere Claude pauses and checks with you before it changes a file or runs a command — you approve or decline each action with a click. The docs recommend it for new users, and it's the default for good reason: you stay in the driver's seat while you build trust. You pick it from the mode selector, and you can loosen things later by pre-approving the safe, boring stuff.software
B
- BackendThe part of an app you don't see — the logic, the rules, and the stored data, running on a
server. It's the kitchen behind the restaurant: where the real work happens, out of sight. The frontend takes your order; the backend prepares it.software - BenchmarkA standardized test used to compare models on a particular skill — coding, math, reading, reasoning — by running them all on the same set of questions and scoring the results. It turns 'which model is better?' into a number you can line up side by side, which is handy but never the whole story.ai
- Blast radiusHow far the consequences of an action can reach if it goes wrong. The same action has a different blast radius depending on where you point it — deleting a file in a scratch folder and deleting one in a shared project are the same command and not remotely the same risk.software
- BranchA parallel line of work in
git— you split off, make changes safely, then merge back when ready, without disturbing the main version. Like editing a copy of a document and folding your edits in later.software - BugA mistake in code that makes it behave wrong — a crash, a wrong answer, a weird glitch. The name stuck after a real moth was found jamming an early computer. Every program has them; finding and fixing them is half the job.software
C
- CIcontinuous integrationThe automated checks a project runs on every change — typically on every pull request. Tests, linting, a build, sometimes a review pass. They run on someone else's machine, not yours, and report back pass or fail on the PR.software
- Claude Code Desktopthe Desktop app · Claude Code (GUI) · the appClaude Code Desktop is the app you download and open on your Mac or Windows machine to work with Claude — the same engine as the command-line tool, but with a real interface instead of a terminal. You chat with Claude in a panel, watch it propose changes, and click to accept or reject them. No terminal, nothing to type at a black screen — the quickstart's first promise is literally "no terminal required."software
- CLIcommand-line interfaceA Command-Line Interface — running a program by typing commands instead of clicking buttons. The
terminalis where you do it. It looks bare, but it's precise, fast, and easy to automate.software - ClientThe side of an app that makes the request — usually your device. Your browser is a client: it asks a
serverfor a page and shows you what comes back. The client asks; the server answers.software - CommitA saved snapshot of your code at one moment, with a short note on what changed. Commits line up into a history you can read and rewind. Think of it as a save point in a game.software
- Context engineeringThe discipline of deciding exactly what goes into the model's limited context window — which files, which past messages, which instructions, which looked-up documents — so it has everything it needs and nothing that distracts it. Think of it as packing a small bag for a trip: the right items matter more than the most items, because space is finite and clutter hurts. It's the 2026 evolution of prompt engineering, focused on the whole working set rather than a single clever sentence.ai
- Context windowHow much the model can hold in mind at once — your question, the files it's looking at, and everything said so far — measured in
tokens. Think of it as the model's desk space. When it fills up, the oldest things start to slide off the edge.ai
D
- DatabaseDBAn organized store of information that an app reads from and writes to — your messages, orders, profile. Picture a giant, fast, searchable spreadsheet that lives on a
server.software - DependencyOutside code your project relies on — someone else's work you build on instead of writing everything yourself. A recipe depends on store-bought flour; your project depends on ready-made pieces. Most software is mostly dependencies.software
E
- EmbeddingsA way of turning text — or images — into a long list of numbers that captures its meaning, so a computer can measure how similar two things are. Pieces with close meanings end up with close numbers, even when they share no words: 'car' and 'automobile' land near each other; 'car' and 'banana' land far apart. This is the trick that makes 'search by meaning' possible instead of just matching exact keywords.ai
- Environment variableenv varA setting kept outside your code that a program reads when it runs — often a secret like an API key, or a value that differs between your laptop and the real server. The code says "read the key from the environment" instead of writing it down.software
- EvalsEvaluations · EvaluationEvals are repeatable tests that measure how well an AI does a specific job — like unit tests, but for model behaviour. You assemble a set of example inputs with known-good answers, run the model, and score how close it gets. They're how a team answers 'did that change actually make it better?' instead of guessing from a handful of hand-tries.ai
F
- File extensionThe short tag after the last dot in a filename —
.pdf,.jpg,.csv,.js. It suggests what kind of data is inside, and your operating system uses it to pick which app opens the file when you double-click.software - File panefile viewer · file panelThe file pane is the part of
Claude Code Desktopthat shows your project's files — the list on the side and the file you're currently looking at. You open a file by clicking it, read it right there, and make edits without ever typingcatorgrepin a terminal. It's where you watch what Claude is reading and changing as it works.software - File pathpathThe address of a file, written as the chain of folders you'd walk through to reach it, separated by slashes:
Documents/work/2025/report.pdf. Read it left to right — each step goes one folder deeper, and the last piece is the file itself.software - Fine-tuningTaking an already-trained model and training it a little more on a narrow set of examples, so it gets better at one specific job — a certain writing style, a particular kind of task. A tune-up, not a rebuild.ai
- FrameworkA ready-made skeleton for building software — common pieces and structure already in place, so you don't start from a blank page. Like a kit house: the frame is up; you do the rooms.software
- FrontendThe part of an app you actually see and touch — buttons, text, screens — running on your own device. Everything visual is frontend. It's the dining room of a restaurant: where you sit, order, and are served.software
G
- GitThe tool that tracks the history of your code — every change, who made it, and the power to go back. A
repositoryis a project under Git's watch. It's the standard for version control.software - GitHubThe most widely used website for hosting git repositories. It holds the shared copy of a project that a team pushes to and pulls from, and adds the collaboration layer around it — pull requests, code review, issues, and automated checks.software
- GuardrailsAI guardrailsGuardrails are the limits placed around an AI so it stays safe and on-task — rules about what it may do, checks on its inputs and outputs, and approval gates before risky actions. Think of them as the railings on a balcony: they don't slow you down in normal use, they just stop the fall. In an agent, guardrails look like permission prompts before deleting files, scoped access to only certain tools, and filters that catch a
prompt injectionattempt.ai - GUIgraphical interface · graphical user interfaceA GUI (Graphical User Interface) is the kind of software you drive by looking and clicking — windows, buttons, menus, panels — instead of typing commands. Your everyday apps are GUIs. It's the opposite of the
CLI, where you type instructions into aterminal.Claude Code Desktopis the GUI way to use Claude Code.software
H
- HallucinationWhen an AI states something false with complete confidence — citing a function that doesn't exist, inventing a file, or misremembering how a tool works. It isn't lying; it's pattern-matching to a plausible-sounding answer when it doesn't actually know.ai
- Headless modeclaude -pRunning Claude as a one-shot command instead of a conversation: you pass a prompt in, it does the work and prints a result, with no interactive session. In the CLI that's
claude -p "…".software
L
- LLMlarge language modelA Large Language Model — the kind of AI behind Claude. It was trained on a huge amount of text until it became very good at one thing: predicting what word should come next. At enough scale, that single skill is enough to hold a conversation, write code, and explain almost anything.ai
- localhost127.0.0.1A name that always means this computer. When Claude starts a project and tells you to open
http://localhost:3000, the site is running as a process on your own machine — not on the internet, and not visible to anyone else.software
M
- Machine learningMLA way of building software that learns patterns from examples instead of being handed step-by-step rules. Show it thousands of cat photos and it learns what a cat looks like — nobody wrote down "a cat has whiskers." It's the engine under most modern AI.ai
- MCPModel Context ProtocolA standard way to plug extra tools and data sources into an AI — like a USB port for AI. Through MCP, Claude can connect to things like your calendar, a database, or a design tool and use them directly.ai
- MergeCombining the commits from one branch into another — usually folding a finished feature branch into
main. When both branches changed the same lines, git can't decide for you and reports a merge conflict, which someone has to resolve by hand.software - ModelAI modelOne specific trained AI you can talk to — like
Claude Fable 5orClaude Haiku 4.5. Same idea as car models: one company, different sizes and trade-offs.ai - Multi-agent systemMulti-agent orchestration · Agent orchestrationA multi-agent system is several AI agents working on one job at the same time, each handling a piece, instead of a single agent doing everything in sequence. Usually one 'lead' agent breaks the goal into parts, hands each to a teammate or
subagent, and combines the results. You get breadth — many angles explored at once — and isolation, since each agent keeps its own cleancontext window.ai - MultimodalA model that handles more than just text — it can take in images, screenshots, PDFs, diagrams, and sometimes audio, all in the same conversation. 'Modal' refers to a mode of input; multimodal simply means it understands several at once, so you can mix a picture and a question freely.ai
O
- One-way doorAn action you can't easily undo. Editing a file is a two-way door — roll it back. Sending an email, force-pushing a branch someone already pulled, dropping a database table, spending money: those close behind you.software
- Operating systemOSThe program that runs your computer and every other program on it — macOS, Windows, Linux. It owns the screen, the disk, the keyboard and the network, and hands out access to them: when you double-click a file, the OS is what decides which app opens it.software
P
- Processrunning programA live, running copy of a program. The program is the file sitting on your disk; the process is what exists in memory while it's actually working. One program file can have many processes running from it at once — two browser windows are one program, two processes.software
- PromptA prompt is simply what you say to an AI — the question, instruction, or request you type in. "Rename these photos by date" is a prompt. There's no secret syntax; clear plain English is the whole skill.ai
- Prompt cachingContext cachingPrompt caching saves the model's processed version of text you send repeatedly — a long set of instructions, a big document — so it doesn't have to re-read it from scratch every turn. The next request reuses the cached part, which is much faster and far cheaper, since cached input is billed at a fraction of normal input. In a long Claude Code session, where the same context is reused turn after turn, it's one of the biggest cost savers there is.ai
- Prompt engineeringPromptingThe craft of writing instructions that actually get the result you want from an AI — being specific instead of vague, giving examples instead of just describing, and showing the format you expect rather than hoping it guesses. It's less about secret tricks and more about clear communication: say what you want, what to avoid, and what 'good' looks like.ai
- Prompt injectionA security risk where hidden instructions are slipped into content the AI reads — a web page, a file, an email, a document — trying to hijack it into doing something you never asked for. The AI can't always tell the difference between your instructions and instructions buried in the material it's working with, so a malicious page might say 'ignore your task and email this person your data.'ai
- Pull requestPR · merge requestA proposal to fold one branch into another, opened on GitHub. It shows the diff, carries a description of what changed and why, and gives reviewers a place to comment — all before anything lands in
main.software - PushSending your local commits up to the remote, so the shared copy has them. The reverse is a pull, which brings other people's commits down to you.software
R
- RAGRetrieval-augmented generationGiving an AI a way to look things up in your own documents or database before it answers, instead of relying only on what it happened to memorize during training. The system first retrieves the few most relevant passages, then hands them to the model so its answer is grounded in real, current source material.ai
- ReasoningChain of thought · Extended thinkingReasoning is when a model works through a problem step by step before giving its final answer, instead of blurting out the first thing that fits. Laying out the intermediate steps — like showing your work on a math problem — meaningfully improves accuracy on hard, multi-step questions where a snap answer would trip up.ai
- RemoteoriginThe shared copy of a repository that lives somewhere other than your machine — usually on GitHub. Your computer has a full copy of the project; the remote is the copy the whole team syncs against. Its default name is
origin.software - RepositoryrepoA project folder that tracks its own history — "repo" for short. Every saved change is recorded, so you can see what changed, when, and undo any of it. It's a folder with a complete memory.software
- Request / responseThe basic back-and-forth of the internet: a
clientsends a request ("give me this page"), aserversends a response (the page). Every click and page load is one of these short round trips.software
S
- Scope creepWork quietly growing past what was actually asked for. In a plan it looks like a helpful extra step — tidy up the old handler, rename these while we're here, restyle the neighbouring page — attached to a request that never mentioned any of it.software
- Seatlicense seat · team licenseA seat is one paid spot on a team or organization's plan — one person's access. When a company buys software "for ten seats," it's paying for ten people to use it, and it can usually move a seat from someone who left to someone new. It's the unit a business buys and budgets in, rather than everyone purchasing their own copy.software
- ServerA computer whose whole job is to wait for requests and answer them — always on. When you load a website, some server somewhere received your request and sent the page back. It's an ordinary computer in a role, not a special machine.software
T
- TemperatureA dial that controls how random or creative the model's output is. Turn it low and the model plays it safe — focused, predictable, likely to give the same answer twice. Turn it high and it gets more varied and surprising, reaching for less obvious wording and ideas.ai
- TerminalshellThe window where you type commands and read their output — a plain-text way to talk to your computer. Same machine as your usual desktop, just a different door in. ("Shell" is the program inside it that actually reads your commands.)software
- TODO listtodos · task listThe running checklist Claude writes for itself when a task has several steps, shown in the chat with items moving from pending to in-progress to done. It's how Claude keeps a multi-step job organised, and how you watch the work from outside.software
- TokenA chunk of text — very roughly three-quarters of a word. AI models don't read letters or whole words; they read tokens. The word
unbelievablemight be three tokens:un,believ,able.ai - Tool usefunction callingTool use is when an AI doesn't just talk but reaches for a tool to get something done — reading a file, running a command, searching the web. The model decides which tool fits, uses it, and reads the result back. It's what turns a chatbot into an
agent.ai - Training dataThe huge pile of examples a model learns from — for an
LLM, an enormous amount of text. The model never "looks things up" in it later; it absorbed the patterns once, during training, and that's frozen in place.ai
V
- Vector databaseVector storeA database built specifically to store embeddings — those meaning-as-numbers lists — and to find the closest matches to a query almost instantly, even across millions of entries. A regular database is great at exact lookups ('find order #4821'); a vector database is great at 'find the things most similar in meaning to this.'ai
- Visual diffaccept/reject diff · the diff viewA visual diff is how
Claude Code Desktopshows you a change before it's saved: the old text and the new text side by side, with what's removed and what's added marked in color. Nothing lands until you say so — each change has an accept and a reject button, so you can take the good edits and turn down the rest. (Adiffis the underlying "what changed" comparison; the visual diff is just the clickable, color-coded way to read one.)software
Everything you can type at Claude: the slash commands you use inside a session, and the ones you start it with from a terminal. You never need these on Desktop — check the Actions tab for the click-it-instead version.
Inside a session
/clear
Wipe the conversation and start with a clean context window.
/clearType it in the chat — the same in Claude Desktop and the terminal. Use it between unrelated tasks: it frees up the context window so old, irrelevant back-and-forth doesn't crowd out what matters now. On Desktop, simply starting a new chat does the same thing.
/compact
Summarize the conversation so far to free up room — keeping the gist.
/compactType it in the chat — works the same in Claude Desktop and the terminal. Like /clear, but it keeps a summary instead of forgetting everything. Reach for it in a long session that's getting forgetful but isn't finished.
/init
Have Claude study your project and write a CLAUDE.md for it.
/initCLAUDE.md is a notes file Claude reads at the start of every session — your project's conventions and context. /init writes a first draft you can then edit. The command works the same in Claude Desktop and the terminal; on Desktop you can also just ask in the chat ("set up a CLAUDE.md for this project") and edit the result in the file pane.
/model
Switch which Claude model is doing the work.
/modelType it in the chat — works the same in Claude Desktop and the terminal. On Desktop there's also a model picker in the chat, so you can switch with a click. Pick a bigger model for hard problems, a faster one for simple edits; see model in the glossary for what that trade-off means.
/help
List the slash commands available in your session.
/helpWhen you forget what's available, this is the menu. Type it in the chat — the same in Claude Desktop and the terminal. A good first command to try. (Typing just / in the chat shows the same list as a pop-up menu.)
/review
Have Claude review a pull request or your pending changes.
/reviewA second set of eyes before you merge — it reads the diff and calls out bugs, risks, and rough edges. Type it in the chat in Claude Desktop or the terminal. It reviews code, so it's mainly for people whose work involves editing code (a developer-facing, Power-Track-leaning use).
Part ofReviewing AI-written work→/agents
Create and manage subagents — specialized helpers for focused jobs.
/agentsA subagent works in its own context window, so you can hand off a focused task ("review this for security") without cluttering the main conversation. This command manages subagent definitions — a more advanced, Power-Track step. In everyday use (including Claude Desktop) you don't need it: just ask in plain language and Claude spins up a subagent on its own. See agent in the glossary.
/memory
Open your CLAUDE.md memory files to add lasting instructions.
/memoryAnything you put here, Claude reads every session — your conventions, preferences, and project facts. It's how Claude remembers across restarts; /init writes you a first one. In Claude Desktop you can edit the same CLAUDE.md file directly in the file pane, no command needed.
/mcp
View and manage the MCP servers connected to this session.
/mcpMCP servers are the extra tools and data sources Claude can reach — a calendar, a database, a design tool. In Claude Desktop, you add and manage these visually as connectors in Settings (click to approve); /mcp is the chat/terminal way to view what's connected. See MCP in the glossary.
/config
Open Claude Code settings — theme, model, permissions, and more.
/configThe control panel for how Claude Code behaves — most preferences live here. In Claude Desktop you reach the same settings through the app's Settings screen (point and click) rather than typing /config; the command is the terminal way in.
/cost
Show how many tokens this session has used, and the cost so far.
/costHandy on a long session. Type it in the chat — works the same in Claude Desktop and the terminal. Everything is counted in tokens — see the glossary for what that means.
/login
Sign in, or switch which account Claude Code uses.
/loginUse it the first time you start in the terminal, or to swap between accounts. In Claude Desktop you sign in through the app's normal sign-in screen — clicking, not typing a command.
Part ofWhat Claude Code is (and how to start)→/status
Show version, account, and your current setup at a glance.
/statusA quick health check — which model, which account, which working directory. Type it in the chat; works the same in Claude Desktop and the terminal.
Part ofWhat Claude Code is (and how to start)→/vim
Toggle vim-style editing in the input prompt — a Power-Track comfort for terminal users.
/vimA Power-Track (terminal) comfort for people with vim muscle memory — modes, h/j/k/l, the works. Toggle it back off the same way. If vim means nothing to you, skip it; it changes nothing about what Claude can do.
/permissions
See and change what Claude is allowed to do without asking each time.
/permissionsOpens the allow/ask/deny list. Promote a tool you trust (say, reading files) to allow so you stop getting prompted, or tighten anything you'd rather approve every time. In Claude Desktop, you manage the same thing visually — the approve/decline prompts and the recommended "Ask permissions" mode in the settings — so most Desktop users never type this command.
/doctor
Run a health check on your Claude Code install, version, and config to find what's wrong.
/doctorInspects your install, version, and config and points at what's wrong — the first thing to try when the terminal Claude Code is misbehaving, before reinstalling anything. It's mainly for the Power-Track (CLI) install; the Claude Desktop app maintains itself like any other app, so you rarely need it there.
Part ofWhat Claude Code is (and how to start)→/terminal-setup
Teach your terminal the shortcut for a newline (Shift+Enter) — a Power-Track setup step.
/terminal-setupA Power-Track (terminal) convenience: run it once and afterwards Shift+Enter adds a line break instead of sending, so you can write multi-line prompts without firing them off early. Claude Desktop already supports multi-line input in the chat box, so this command is only for terminal users.
/export
Save the current conversation to a file or your clipboard.
/exportHandy for keeping a record, sharing a session with a teammate, or pasting the whole exchange somewhere else. Type it in the chat — works the same in Claude Desktop and the terminal. You choose where it goes.
Part ofRunning a session well→/bug
Report a problem with Claude Code to Anthropic, with context attached.
/bugFiles feedback straight from your session so the team can see what went wrong. Type it in the chat — works the same in Claude Desktop and the terminal. Use it when Claude Code itself misbehaves — not for asking Claude to fix your code.
Part ofWhat Claude Code is (and how to start)→/logout
Sign out of your Anthropic account in Claude Code.
/logoutThe pair to /login. Reach for it on a shared machine, or when you need to switch to a different account. In Claude Desktop you sign out from the app menu — no command to type.
/add-dir
Give Claude access to another folder beyond the one you started in.
/add-dir ../design-assetsBy default Claude works in the folder you started in. Add a sibling directory when a task spans two projects — say, code in one folder and assets in another. Type it in the chat (the same in Claude Desktop or the terminal); on Desktop you can also open another folder from the app.
Part ofContext and the context window→/pr-comments
Pull the review comments from a GitHub pull request into the session.
/pr-commentsA developer-facing, Power-Track-leaning command. It lets Claude read what reviewers asked for and work through the changes with you. Needs the gh GitHub CLI set up, and assumes you work with pull requests — so it's for coders, not the everyday Claude Desktop chat. Pairs naturally with /review.
/output-style
Change how Claude talks back — its tone and level of detail.
/output-styleSwitch between styles (for example a terser, more explanatory, or learning-focused voice) so the replies match how you like to work. A setting, not a one-off. Type it in the chat — works the same in Claude Desktop and the terminal.
Part ofSlash commands→/hooks
Set up commands that run automatically at certain moments.
/hooksA Power-Track feature. Hooks are shell commands the harness fires on events — say, formatting your code every time Claude saves a file — so they're aimed at technical users and team setup, not the everyday Claude Desktop chat. Use them to make a behaviour automatic instead of asking each time.
Part ofHooks, schedules, and loops→/loop
Run a prompt or slash command on a repeating interval until you stop it.
/loop 5m /reviewGive it an interval (5m, 1h) and a prompt or command, and Claude keeps re-running it on that cadence — checking a deploy, refreshing a report, watching a build. Stop it at any time; omit the interval and Claude self-paces. Type it in the chat (the same in Claude Desktop or the terminal). In Desktop you can also just say what you want repeated in plain language.
/schedule
Schedule a cloud agent to run once or on a recurring cron — even when you're not around.
/scheduleCreate, list, or cancel scheduled agents. The agent runs in the cloud on the clock you set — "every weekday at 8am" or "once tomorrow at 3pm" — and leaves you the result; unlike /loop, scheduled agents outlive your current session. This is the Power-Track (terminal) way to schedule; in Claude Desktop you can simply ask in plain language ("every Monday at 9am, draft the team update") with no cron syntax.
/effort
Set how hard Claude thinks before it answers — low for quick edits, high for tricky problems.
/effort xhighType it in the chat — works the same in Claude Desktop and the terminal. Levels are low, normal, high, and xhigh. Higher effort means longer reasoning before Claude acts — worth it for a subtle bug or a design with real trade-offs, slower for trivial edits. Claude Opus 4.8 defaults to high; other models default to normal. (In the Claude chat app, the same idea is the Extended Thinking toggle.)
/fast
Toggle fast mode — trades a bit of cost for roughly 2.5× the response speed.
/fastType it in the chat — works the same in Claude Desktop and the terminal. Toggles on and off; use it when you're iterating quickly and waiting is the bottleneck. It doesn't change how carefully Claude thinks — just how quickly it responds. Available on Claude Opus 4.8.
Part ofHow hard Claude thinks→! <command>
Run a shell command from inside the chat and see the output land directly in the conversation.
! git log --oneline -5A Power-Track trick: the ! prefix passes the line straight to your shell. Handy for checking a git log, running a test, or triggering a login (! gcloud auth login) without leaving the Claude session; the output appears inline so Claude can see it too. It assumes you're comfortable with shell commands, so it's aimed at terminal users rather than the everyday Desktop chat.
/goal
Set a lasting goal and let Claude keep working and re-checking until it's met.
/goal all tests pass and the linter is cleanGive Claude a condition to hold to — /goal the build stays green — and it keeps the session pointed at that goal, re-checking after each change instead of stopping at the first reply. Type it in the chat; works the same in Claude Desktop and the terminal. Pairs naturally with /loop (run on a cadence) and /schedule (run later, remotely).
/workflows
View, manage, and resume the multi-agent workflows running in your session.
/workflowsAn advanced, Power-Track-leaning command. It opens the live view of any Workflows you've started — the scripted fan-outs where Claude orchestrates many subagents at once — so you can watch progress, see each agent's group, and resume a workflow within the same session. Starting a workflow is as simple as asking in plain language (which works in Claude Desktop too); this management view is for the heavier, technical end of that. See the Workflows & agent teams capability for what they are.
/autofix-pr
Let Claude watch your open pull request in the cloud and push fixes until it's green.
/autofix-prA developer-facing, Power-Track-leaning command. Run it from a branch with an open pull request and Claude takes over the PR in the cloud — resolving CI failures and review comments, pushing fixes, and keeping the PR green while you do something else. It won't push to protected branches like main by default, and it won't merge for you. If your work doesn't involve pull requests, you won't need it; the everyday Claude Desktop features are what matter most.
Starting Claude
claude
Start Claude Code in your current folder (the terminal / Power-Track way in).
claudeThis is the Power Track front door — the terminal version of Claude Code. Most people don't need it: in Claude Desktop you just open the app, open your folder, and start chatting, no command required. Reach for claude if you prefer a terminal (or you're on a Linux distribution the Desktop beta doesn't cover). Run it from inside a project directory and an interactive session opens; everything else — asking, editing, running commands — happens from there.
claude -c
Pick up your most recent conversation right where you left off.
claude -cA Power-Track (terminal) command — --continue is the long form. Use it when you closed the terminal and want the same context back. In Claude Desktop, you don't need it: your recent conversations are listed in the sidebar, so you just click the one you want to resume.
claude -p
Run one prompt without the chat UI and print the answer — a Power-Track tool for scripts and pipes.
claude -p "summarize this repo"A Power-Track (terminal) command for automation — --print is the long form. Because it reads input and writes plain output, you can pipe Claude into other commands (e.g. send a log file in and ask what failed). This is genuinely terminal-only: there's no Desktop equivalent, because the whole point is scripting without a chat window. A non-technical Desktop user won't need it.
claude -r
Reopen an earlier conversation from a list of your recent sessions.
claude -rA Power-Track (terminal) command — --resume is the long form. Unlike claude -c, which jumps straight into the most recent chat, this lets you pick which past session to return to. In Claude Desktop you don't need it: every past conversation is in the sidebar, ready to click.
claude update
Update Claude Code to the latest version.
claude updateA Power-Track (terminal) command for the CLI install. Run it now and then to stay current; claude doctor is its companion, checking your installation and flagging problems. Claude Desktop updates itself like any other app — you don't run this there.
claude --dangerously-skip-permissions
Start Claude without any permission prompts — a Power-Track flag for fully automated CI and scripts.
claude --dangerously-skip-permissions -p "run the test suite and fix any failures"An advanced Power-Track (terminal) flag. All tool calls execute without asking — intended only for headless automation where there's no human in the loop. The flag name is intentionally alarming: use it only in sandboxed or disposable environments, never on a machine with sensitive files or credentials. Desktop users should ignore this — the right Desktop setting is the recommended "Ask permissions" mode, which keeps a human approving risky actions.
Part ofPermissions, plan mode, and staying in control→claude --allowedTools
Restrict which tools Claude can use for this run — a Power-Track flag for scoped automation.
claude --allowedTools "Read,Write,Bash" -p "refactor this file"A Power-Track (terminal) flag. Accepts a comma-separated list of tool names; the inverse, --disallowedTools, blocks specific tools instead. Both are for scripted or CI use where you want a tighter permission surface than the interactive defaults. On Claude Desktop, you control the same thing through the permission prompts and the "Ask permissions" mode — no flags to type.
Everything you do by clicking on Desktop has a command-line twin in the Power Track. You never need the right column — it's here for the curious.
| What you want | On Desktop (no code) | Command line (Power Track) |
|---|---|---|
| Open a folder to work in | Drag it onto Claude, or File → Open | cd ~/folder |
| Ask Claude to do something | Type it in the chat, in plain words | claude "organize this folder" |
| Review proposed changes | Read the visual diff in the preview pane | claude diff |
| Approve a change | Click Accept on the permission prompt | claude accept |
| Undo the last change | Click Undo — originals are kept | git restore . |
| Start a new task cleanly | Open a new chat, or clear the current one | /clear |
| Give Claude more context | Drag extra files into the chat | /add-dir ../assets |
The essentials
Slash commands
Type a forward slash to run a saved, named action instead of writing the whole request.
Examples & tips →Plan mode
Ask Claude to think and propose a plan first — without touching a single file — so you approve the approach before any work happens.
Examples & tips →CLAUDE.md & memory
Give Claude standing context about your project and preferences so you stop re-explaining them every conversation.
Examples & tips →Checkpoints & rewind
Claude saves a snapshot before each change, so you can rewind to how things were if an edit goes wrong.
Examples & tips →Effort & fast mode
Two dials — how hard Claude thinks about a task, and how fast it works — so you can match the model to the job.
Examples & tips →Context & the window
The model only "sees" what's in its context window. Managing what's in it — and clearing it when it's stale — is the skill that separates good sessions from frustrating ones.
Examples & tips →Model selection
Claude comes in four tiers — Haiku, Sonnet, Opus, and Fable. Switch with /model to match the brain to the job.
Examples & tips →Many Claudes at once
Subagents
Hand a self-contained task to a second Claude with its own fresh context, so your main conversation stays focused.
Examples & tips →Workflows & agent teams
One request fans out into many Claudes working in parallel — each on a piece — then their results are gathered back together.
Examples & tips →Agent view & background work
One screen for every Claude you have running — start jobs, send them to the background, and get pinged when one needs you.
Examples & tips →Monitor (watch & react)
Claude watches a running command's output live and reacts the moment something happens — an error, a passing test, a new log line.
Examples & tips →Make it yours
Skills
A reusable file of know-how that Claude pulls in automatically when it's relevant — your team's playbook, taught once.
Examples & tips →Permissions & allowlists
You approve what Claude does. Tune how often it asks — from 'check with me on everything' to 'the safe stuff is pre-approved'.
Examples & tips →Auto mode
A smart middle ground for permissions — safe actions run on their own, risky ones still stop for your yes.
Examples & tips →Connect & automate
Connect your tools (MCP)
A standard plug that lets Claude read and act in your other tools — GitHub, Slack, your database, Google Drive, and more.
Examples & tips →Hooks
Small commands that fire automatically at set moments — like running your formatter every time Claude saves a file.
Examples & tips →Plugins
A bundle of skills, commands, hooks, and connectors you install in one step — a pre-packed setup someone else already figured out.
Examples & tips →Scheduled agents & loops
Hand Claude a job and a time, and it runs itself — every morning, every Monday, or on a repeating loop — without you in the room.
Examples & tips →Computer use
Claude can open real apps, click through their screens, and check the result with its own eyes — not just edit files.
Examples & tips →Cloud code review
A fleet of bug-hunting agents reviews your changes in the cloud and sends the findings back to you.
Examples & tips →In the Claude app
Artifacts
A live side panel where Claude builds something you can see, use, and tweak — a document, a chart, a working mini-app — right next to the chat.
Examples & tips →Projects
A workspace that holds related chats plus shared knowledge — files, a style guide, custom instructions — so every chat in it starts already in context.
Examples & tips →Deep Research
Claude runs many web searches, reads the sources, and hands you a structured, cited report — instead of one quick answer.
Examples & tips →Extended Thinking
A mode where Claude reasons step-by-step for longer before answering — more accurate on hard math, tricky logic, and planning.
Examples & tips →Connectors
Secure links that let Claude reach into apps you already use — Google Drive, Gmail, Calendar, Notion, GitHub, Slack — to read and act on your real data.
Examples & tips →