The Next Mountain · Resources

Claude, and everything in its toolbox

Most people meet Claude as a chat box and never find the rest. This is the whole toolbox in one page: what each part is for, when to reach for it, and what it can and cannot see. Start with what you are trying to end up with — the tool follows from that, not the other way round.

Current as of 28 Aug 2026Version 1.0Maintained by The Next Mountain

Start here

Pick what you are trying to produce — it opens the right card below

What do you want to end up with?

Four places to work

Different rooms in the same building — you pick by the deliverable

Chat

One conversation, you steering every turn. Where most people spend most of their time, and rightly so.

Use it when: You are thinking something through, drafting from a rough start, or rewriting to a standard — anything where you want to see the work and push back on it.
You get: An answer in the thread, plus artifacts: documents you can open, edit, and share on their own.
Its limit: It waits for you. Nothing happens between your messages, and it starts fresh each conversation unless the work sits in a project.

What it can see

By default: — Only what is in the conversation — what you typed, pasted, or attached.
If you enable it: — Whatever connectors you have switched on for that chat.
You control: — Every attachment and every connector, per conversation.

Cowork

Claude works across files and connected apps over many steps, then hands back the finished thing.

Use it when: The job has many steps and you do not need to watch each one. You are delegating an outcome rather than having a conversation.
You get: A finished file, report, or briefing, produced in its own session that you review when it is done.
Its limit: It cannot stop to ask a question mid-run. Anything vague in your request gets decided without you.

What it can see

By default: — Your connectors, your installed skills and plugins, and files saved to your Claude account.
Approval mode: — You set how much it may do unattended. Anything that writes, sends, or posts should require sign-off until you trust it.
You control: — The approval mode, and whether it works in a folder on your computer or only in the cloud.

Design

A visual canvas with a chat panel beside it.

Use it when: The deliverable is something people look at rather than read — a deck, a one-pager, a clickable prototype.
You get: A canvas you refine by conversation, inline comment, or direct edit. Exports to PowerPoint, PDF, or Canva.
Its limit: A separate product with its own page, not a feature inside chat.

What it can see

By default: — What you describe and upload.
If you allow it: — Your organization's design system, so output matches the house style.

Code

An agent that works inside a code repository.

Use it when: The deliverable is working software. Runs in a terminal, a desktop app, or the web.
You get: Commits, pull requests, and scripts.
Its limit: Built for developers. Listed here so the map is complete and the name is familiar when it comes up.

What it can see

By default: — The repositories you give it and the connectors you attach.
Guardrail: — Its unattended jobs push only to their own branches unless you deliberately widen that.

Three things you install once

Set up in the directory, then available in the places above

Connectors

The reach into your mail, calendar, files, and other systems.

What it is: A pre-built link to a service — mail, drive, chat, a database — that you authorize once.
Use it when: The work needs your real records rather than what you can paste in.
Note: The highest-stakes item on this page, and the one worth setting up slowly.

What it can see

Boundary: — Claude acts as you, with your permissions. If you cannot see a document, neither can it.
Scope: — Each connector can be narrowed — for example, allowed to search and summarize mail but not to send it.
Who controls it: — You authorize it; on a company plan, an administrator can restrict which connectors are available at all.

Skills

Your written instructions for one specific task.

What it is: A short document telling Claude how you want a particular job done. It loads only when your request matches it.
Use it when: You do the same thing more than twice and care how it turns out.
Note: A skill can tell Claude to use a connector. That pairing — a procedure plus reach — is where most of the value sits.

What it can see

On its own: — Nothing. A skill is instructions, not access.
Who controls it: — Whoever owns the skill. On a company plan, skills can be provisioned centrally so everyone works to the same version.

Plugins

Skills, shortcuts, and the connectors they need, bundled as one install.

What it is: A packaged toolkit for a job function. Open the box and everything inside is already wired together.
Use it when: A whole workflow should arrive identically for everyone rather than being assembled person by person.
Note: The unit of distribution. This is how a team gets set up the same way in one step.

What it can see

Boundary: — Installing a plugin also grants whatever connectors are bundled inside it — so read what is in the box.
Who controls it: — You, for your own account. On a company plan, an administrator can approve or push plugins to the whole team.

Where work is kept, and what comes out

The two most commonly missed items in the toolbox

Projects

A container that holds related conversations with the files and standing instructions they share.

Use it when: The work will run for weeks and across more than one conversation.
You get: Context that persists. Reference documents and your standing instructions stay loaded, so you stop re-explaining the background every time.
Note: If you are pasting the same background into a fourth conversation, that is the signal to make a project.

What it can see

Contents: — Only the files and instructions you put in it.
Sharing: — Private by default. Sharing a project shares its files and instructions with everyone you add.
Boundary: — Connectors work in private projects only, and conversations that pulled in synced content cannot be shared.

Artifacts

The standalone thing a conversation produces — a document, a working page, a diagram.

Use it when: You need the output as an object you can keep, not as a message you scroll back to find.
You get: Something you can open beside the chat, edit through further instructions, and share on its own.
Note: Claude often makes one without being asked, whenever the answer is a thing rather than a reply.

What it can see

Sharing: — Sharing an artifact is a separate decision from sharing the conversation that made it — check which one you are handing over.

Two ways to make it repeat

Same idea, two surfaces, two names

Scheduled tasks

A Cowork job that runs on a cadence you set — hourly, daily, weekly, or on demand.

Use it when: The same check or summary needs to happen whether or not anyone remembers it.
You get: A finished output waiting for you, produced in its own session. It runs remotely, so your computer can be closed.
Its limit: It cannot ask you anything. Whatever you wrote when you set it up is what runs, at the quality you set, every time.

What it can see

By default: — Your connectors, skills, and plugins, exactly as a normal Cowork session would.
Approval mode: — Set when you create it, and it holds for every future run.
Boundary: — A task that needs files on your own computer can only run while that computer is on.

Routines

The same idea inside Claude Code: a saved job with a repository attached.

Use it when: The recurring work is development work — nightly checks, automatic review of incoming changes.
You get: A job that fires on a schedule, from a direct call, or in response to an event in the code repository.
Note: If you are not writing software, this is the one item here you can safely ignore. It is on the page so the vocabulary is not confusing when someone mentions it.

What it can see

By default: — The repositories and connectors attached when it was created.
Guardrail: — Pushes to its own branches unless that restriction is deliberately lifted.

Two dials, set every time

Independent of each other — the strongest model does not automatically think the longest

Model

Which Claude answers you. The picker sits under the message box.

HaikuSonnetOpus

Haiku is fast and light. Sonnet is the everyday workhorse. Opus is for the hard ones.

Effort

How long it thinks before it starts answering. Sits beside the model picker.

LowMediumHighMax

If you are skimming the reply for the useful part, the dial is set too high for that task.

Five things worth knowing before you start

Claude sees only what you hand it. A conversation has no access to your systems until you switch a connector on, and a connector never sees more than you can see yourself.
Installing something grants what is inside it. A plugin can carry several connectors. Read the contents before you install, the same way you would with any other software.
Anything unattended runs at the setting you gave it. A scheduled job cannot ask a question, so a vague instruction produces a quietly wrong result every week until someone reads it.
Sharing is per object. Sharing a project, a conversation, and an artifact are three separate decisions. Check which one you are actually sending.
The tool changes underneath you. Several items on this page did not exist a year ago. Check the date at the top before you trust it.

Keep this page honest

This page has a shelf life. Compass is how you hear when it moves.

Half of what is on this page did not exist a year ago, and some of it will be wrong by Christmas. Compass is our free weekly email: what actually shifted in AI, and what it means for the people you lead. One story, one practice, one ask. Every Monday, about five minutes.

When this guide changes, Compass readers hear about it first.

Already reading Compass and want the room behind it? Base Camp is where we put this to work together, monthly.

Claude orientation · The Next Mountain · Version 1.0 · 28 Aug 2026Out of date, or want help putting it to work? Book a conversation

Claude is a product of Anthropic. This guide is written and maintained by The Next Mountain for the leaders we work with; we are not affiliated with Anthropic, and product names are used for identification only.