Claude and ChatGPT 2.0: The Tour

This is the detailed tour behind the book's first page, Claude and ChatGPT 2.0. That page explains what changed and what stays your responsibility. This one shows each feature, compares the two apps, and advises which model to use for which job.

You do not need to read it all now, or remember it. Skim it to see what the apps can do, and come back when a chapter uses a feature. Part 0 walks you through setting up the apps step by step. Words like repository, worktree and pull request belong to software work. You can skip any paragraph that uses them.

This book uses the two desktop apps as its default platforms, because they can do the most for the work it teaches: they work on the files in your own folders, use a built-in browser, control other apps on your computer, and run the coding agents you meet in Part III. The web and phone apps can do much of the same, and the book points out where they differ.

Claude 2.0

As verified on November 5, 2026.

Two places to work

The Claude desktop app has two main places to work.

Conversations. One conversation now does everything. You no longer choose between chatting and handing off a task. You describe what you want, and Claude decides whether a quick answer is enough or whether the request needs real work. Anthropic used to call that working mode Cowork. It now happens inside any conversation, and this book calls it simply a task.¹,² A quick question gets a quick answer. A request for research, a report, a spreadsheet or a presentation becomes a task that searches, reads files, runs code, and returns editable files in the same conversation. Several tasks can run at once.

In a conversation Claude can:

• Work with your files. It can read, organize and create files in a folder on your computer that you grant it.
• Use a browser. It can open pages, read them, click and fill in forms, in the app's built-in browser or in your own Chrome through Claude in Chrome.
• Use your computer. With computer use, it can click, type and move between apps on your screen.³
• Use your apps. Connectors to apps such as Google Drive, Gmail, Microsoft 365 and Slack work during tasks. Skills teach Claude how you want a job done, and plugins bundle skills and connectors for a role.
• Draw in the conversation. Charts, diagrams and small interactive visuals appear inside the reply.
• Keep going without you. Tasks continue in the cloud if you close your laptop, and recurring tasks can run in the cloud on a schedule. Work that needs your local files, the built-in browser or computer use needs the desktop app open.
• Follow you to your phone. Tasks started on the desktop appear in the mobile app, where you can answer questions, redirect the work or start something new.

One setting governs how freely Claude acts in the whole conversation. Auto lets Claude work without asking at each step, with automated safety checks before each action. Manual asks before each action. You can change it at any time, and you can stop or redirect Claude while it works. By default Claude asks before permanently deleting files.⁴,⁵

Because Claude now picks the route itself, this setting carries real weight. One request can start a task that uses a connector, then a browser, then your screen, without a separate step where you choose to hand the work over. If you want to see each action before it happens, choose Manual. Chapter 7, The Authority Envelope, teaches how to match each permission to what a worker may do.

Code. The Code tab is Claude Code inside the desktop app. Despite its name, it is not only for programmers. You talk to it in plain language, and many people who use it never read the code it writes. Many of the newest features appear here, and most of the rest of this section describes it. Part III teaches it properly.

New artifacts: documents, slides, designs, dashboards and motion

An artifact is a piece of finished work that gets its own window beside the conversation: something you keep, edit, present or share, not just read once. Artifacts work in conversations and in the Code tab. Five new kinds join the familiar pages, diagrams and small tools:

Docs, Slides and Design are the most widely available. Dashboards and Motion are in beta on higher plans.⁶,⁷

Three properties make artifacts more than files:

• They start private. You decide whom to share them with.
• They can be live. On paid plans an artifact can read from and write to connected apps, and keep its own data between visits. Each person who opens it connects their own apps.
• They can think. An artifact can call Claude itself, so a shared tool can answer questions or coach its users. Each person's use counts against that person's own plan, not the creator's. The From Skills to Income page shows why that matters if you want to sell what you build.

The Code tab's new layout

The Code tab was rebuilt so you can see what Claude is doing without reading its transcript.

The session rail is a column of cards beside the conversation, which you can show or hide. Environment shows where the session runs (on your computer, in the cloud, or on a remote machine), which folder it works in, and a switch for Remote Control. Context shows how full the session's context is: everything Claude holds in mind at once, meaning your messages plus the connectors, skills, agents and memory files it has loaded. When it fills, older detail is compacted into a summary, so the card helps you spot a connector or memory file crowding out your work. Changes lists what the session has changed, so you can open each change and comment on it.⁸

Panes open beside the conversation: a terminal, the list of changes, a browser that previews what Claude is building, the files it is working with, and its plan and background tasks. You may never open most of them. They are there for the moment you want to see for yourself.

Sections group the sidebar's sessions by project, client or topic, under names you choose.

Asking without interrupting

Claude often works for minutes at a time, and you will have questions while it does.

• Side chat. Open one with /btw or a keyboard shortcut. It can read the main conversation up to that point and answer you, but nothing you say there is added back. Side chats are not saved after you close the app.
• Fork. Copy the conversation from a chosen point into a new session, to try a different direction without disturbing the original.
• Rewind (reported). Go back to an earlier point in the session with /rewind, like undo.⁹
• Split view. Open sessions side by side, from a session's Open in menu or by Option-clicking (Alt-clicking on Windows) it in the sidebar.

Keeping the work running when you step away

• Remote Control connects a session on your computer to the Claude mobile app and the web, so you can read it, answer its questions and approve its requests from your phone. Turn it on per session from the session rail, or in Settings turn on Connect new sessions to Remote Control.
• Keep computer awake stops your computer from going to sleep from inactivity while the app runs your work. Closing the laptop lid still puts it to sleep.
• Cloud sessions run on Anthropic's computers, so they continue with your laptop closed. A cloud session in the Code tab cannot see the files on your computer. It works from what you give it, such as a GitHub repository or uploaded files.
• Scheduled tasks start a session at a set time. A local task needs your computer on. A cloud task does not, and can also start on a GitHub event or an API call.

These pair well with /goal, which keeps Claude working until a condition you set is met. You state the finish line. In the weekly supplier payment run at Brightline Wholesale Supply, the company whose bill payments you follow through this book, it might be "all 15 invoices have a pay, hold or ask decision, and the run total matches the checking program". After each turn, a separate small model reads the conversation and judges whether the condition holds. If not, Claude starts another turn instead of stopping to ask you. The goal also ends if the checker judges the condition impossible, or if a turn fails on an error you have to fix, such as a used-up credit balance. You can clear a goal yourself at any time with /goal clear.¹⁰

Two cautions. Write a finish line Claude's own output can show, never "make it great". And the separate model only reads what the conversation shows. It does not open files or rerun anything. It is a second opinion, not an independent check. A met goal is evidence that the work is finished, and it never makes the next action permitted.

Read the figure from left to right, then follow the curved arrow back for each turn that falls short. On the right are the two ways the checker ends a goal: met, or judged impossible. The note in the middle is the third way, an error you must fix. The two notes along the bottom are the cautions above.

Where the work actually runs

It helps to know what happens behind the screen when Claude "keeps working in the cloud".¹¹,¹²

• A fresh virtual machine for each session. A cloud session runs on a virtual machine: a complete computer, simulated in software, on Anthropic's servers. Each session gets its own, created when the session starts and deleted when it ends. Sessions never share one, even inside the same company.
• Idle sessions lose their machine. In the Code tab, if a cloud session sits idle long enough, Anthropic takes its virtual machine back. When you reopen the session, Claude starts a fresh one. Your conversation comes back. Work that was still running in the background does not, and Claude tells you what stopped so you can rerun it.
• What survives and what does not. The conversation and the files saved to your account survive. Anything Claude installed or left on the virtual machine does not. So ask for results as files you keep.
• What it can reach. Only websites on an allowed list, through a filter the session cannot change, and nothing inside your company's own network. Enterprise organizations start with no internet access at all. Your keys for GitHub and your connectors stay outside the machine, and Anthropic's servers add them when a request leaves it.
• Your own files. A cloud session reaches your computer only through the desktop app, and only folders you have connected. Each request is checked against your permissions. Files read this way are processed on Anthropic's servers, not only on your computer.
• Local sessions. A Code tab session set to Local runs on your own computer, in the folder you chose. Older tasks started on the desktop before October 2026 ran inside a virtual machine on your own computer instead.

Anthropic documents a new machine per session. It does not say whether each new task inside the same conversation starts a new session, so do not count on anything carrying over between tasks except your conversation and your files.

Three ways to run several things at once

Nested sessions (reported) split one job into side jobs. They are described in walkthroughs of the new app but were not yet in Anthropic's documentation when this page was verified. Claude can start sessions of its own that run in parallel, appear under the main session, and report back to it. Each is a full session you can open and talk to directly: you might have five competitors researched at once, then ask about one of them without cluttering the main conversation. Ask for it ("split this into a separate session for each competitor"), or start one from a card in which Claude suggests related work.¹³,¹⁴,⁵

A coordinating session manages sessions already running. This is a way of working, not a button. Every Code session can list, read, message, rename and archive your other sessions, so you pin one to the top of the sidebar and use it as a control center. Ask it "what's waiting on me?" and it tells you which sessions have finished, which need approval and which are still working. It sees your 20 most recently active sessions on this computer, not cloud sessions, unless you ask it to look further.¹⁵

Projects carry ongoing work with shared memory. A project is one ongoing conversation in which Claude acts as a coordinator: you tell it what needs doing, and it starts a thread for each task. Threads run in the cloud by default, so they keep going with your laptop closed, though you can ask for one to run on your computer. Every new thread starts from the project's instructions, and cloud threads also read the project's memory, so a rule you state once reaches all of them. An Overview pane shows what is waiting on you, and a Library holds the files you added and the files threads produced.¹⁶ These are not the projects in Claude conversations, which hold files and instructions for a set of conversations. Same word, different feature.

The arrows show who starts the work. In nested sessions the main session starts its side jobs and they report back. A coordinator starts nothing. It watches and steers sessions that already exist. A project starts cloud threads that share its instructions and memory.

Parallel work uses your plan's limits faster, and so does standing context: connectors, skills and memory files load into every new session, and each side-chat question rereads the main conversation. Start with two sessions and add more once you trust the results.

Mods: changing the app itself

Mods change how Claude Code looks and behaves. Describe the change, for example "make me a mod that shows my weekly usage above the message box", and Claude writes the mod, installs it, and reloads it in your session. A mod can add a pane or a status line, show a notification, rewrite or block a tool call, answer a permission request, or hide secrets from the screen. Anthropic's examples include a pane that follows your builds, a confirmation step before any command touches production settings, and an audit log of what other mods do.¹⁷

Mods are not sandboxed: they have the same access to your computer as Claude Code itself. Install only mods you have read or that come from a source you trust. In companies, administrators decide which plugin sources are allowed, and a built-in safety mod loads first to block the riskiest changes.

Briefing Claude 2.0

The new app needs less instruction than most people still give it. Anthropic's guidance for Opus 5.5 is to start at medium effort, the default, and test other levels on your own work. It suggests stating the condition for done up front, and removing instructions that tell Claude to think carefully, because the model decides how much to think and the effort setting is the main control.¹⁸ Panaversity's own research adds two habits: give the whole task in one message, and leave out lines such as "double-check everything". A specific condition for done adds far more.

Chapter 5 teaches this as the Four-Part Brief: the outcome you want, the format of the result, the inputs to use, and how much autonomy the worker has. The payment-run brief on the first page shows all four. The new apps shift its emphasis. The outcome carries a done-when Claude can check itself. Every brief still covers all four parts, but inputs, format and autonomy can be short when something else already settles them, such as project instructions, approved sources and enforced permissions. End with "Show what you checked and what the checks found, and list anything you could not verify", so that you review evidence, not a claim. When you are not sure what you want, ask Claude to interview you until it could write the brief itself, then start a fresh session from that brief.

ChatGPT 2.0

As verified on November 5, 2026.

One app for Chat, Work and Codex

OpenAI's new desktop app for Mac and Windows brings three experiences into one app, on every plan, including Free.¹⁹,²⁰

• Chat is everyday conversation.
• Work is an agent for longer jobs. It gathers information from your apps and files, splits a project into steps, works for hours if needed, and returns finished materials.
• Codex is OpenAI's coding agent, with its own view and history.

You choose Chat or Work with a toggle above the message box, and switch to Codex separately. Chat and Work conversations share one list of recent chats, and projects hold related files and context for both.²¹,²²

GPT-6 Astra, OpenAI's most capable model, is (reported) the model behind Work on Plus and higher plans. In Codex you can choose among the GPT-6 models. Everyday Chat runs on lighter GPT-6 models. Which models you see, and which is the default, depends on your plan and your workspace settings, so check the model picker in your own app. Chat answers can now mix text with interactive charts, forms and small tools when that suits the question.²³,²⁴,²⁵,²⁶

What Work can do

• Work with your files and apps. On the desktop it can use local files and apps. Work you start in the cloud continues on the web and on your phone.
• Use a browser. The desktop app has a built-in browser for research and web tools. A cloud browser runs on a separate remote computer with its own sign-ins, never your own browser's passwords or tabs. ChatGPT in Chrome adds ChatGPT to Chrome's sidebar, and sites can offer ChatGPT their own tools directly through WebMCP.²⁷,²⁸,²⁹
• Use your computer. Computer use lets ChatGPT click, type and move files across your apps in the background, once or as part of a scheduled task. Record & Replay turns a workflow you demonstrate into a reusable skill.
• Produce finished work. Documents, spreadsheets, presentations and web apps, often built to your templates and reference files. Sites, in beta, are interactive web apps such as dashboards, trackers and internal portals that you can share with your team or publicly, and that update as their data changes.
• Build decks with your team. Slides turns a conversation, a file, an outline or a template into a presentation, with speaker notes and sources, in your brand style if you supply it. Ask "create a slide deck from this chat" and the deck opens in a slide editor where you can change text, charts, tables and images yourself. Decks live in Space, ChatGPT's shared area for pages and decks. You invite people to view or edit, and teammates can select part of a slide and ask ChatGPT to change it. You can present from ChatGPT, or export a copy to PowerPoint, PDF or Google Slides. The live deck stays in Space.³⁰,³¹
• Use your apps. Plugins connect ChatGPT to tools such as Slack, Teams, Google Drive, SharePoint, email, calendars and CRMs. ChatGPT picks relevant plugins itself, or you name one by typing "@" and its name.
• Repeat what worked (reported). Skills package a way of doing a job so ChatGPT follows the same method each time. The results still need checking. When you like how ChatGPT handled something, ask it to turn that into a skill. You can also describe one to the skill creator, write one in an editor, or upload a skill file. You call a skill by typing "@" and its name, in Chat or in Work.³²,³³,³⁴,¹⁴
• Run on a schedule or on an event. Scheduled tasks run once, on a schedule, or to watch for a change. They can also start when something happens: a new Gmail message, a message in a Slack channel, or activity on a GitHub pull request. Plugin builders can offer their own events through MCP.³⁵,³⁶,³⁷
• Remember. Memory saves useful preferences and recurring workflows when you turn it on. Opt-in features can also build memories from what you do on screen.
• Keep some chats private (reported). You can lock a chat behind a four-digit PIN, or Face ID, Touch ID or device unlock on a phone. A locked chat disappears from your history and search and is found under Projects > Locked chats. Locking hides a chat from people using your device. It is not encryption, and a forgotten PIN can only be reset by deleting every locked chat.³⁸,³⁹

Where the work actually runs

ChatGPT's cloud work also runs on separate, protected computers, but OpenAI describes a different design from Claude's.⁴⁰,⁴¹

• Machines that can be reused. Work's cloud tasks run in sandboxes backed by virtual machines on OpenAI's servers. OpenAI says Work can reuse an environment across tasks, or replace it while keeping the state it needs, so not every task gets a new one. Saved state is what lets Work resume an interrupted task and refer back to earlier work.
• Deleting is not immediate. Ending a task or deleting a chat does not immediately remove everything stored with it. Your workspace's retention settings decide how long saved state is kept.
• What it cannot reach. The cloud environment does not inherit your local files, installed apps, browser sign-ins or company network. The cloud browser is another separate computer with its own sign-ins.
• Steps on your own computer. A Work task can run some steps on your connected computer, which must be online. If it is not, an existing task can continue in the cloud, but without your files.
• Local work in the desktop app and Codex runs on your own computer, inside your operating system's own sandbox (Seatbelt on a Mac, bubblewrap on Linux, the Windows sandbox on Windows), not in a virtual machine. Your permission mode sets what that sandbox allows.

Codex in the same app

Codex is to ChatGPT what the Code tab is to Claude. Like the Code tab, it is not only for programmers: OpenAI says you can use it for research, documents, presentations and other knowledge work.⁴² What it adds is developer views and technical detail, which matter most when software gets built, reviewed and tested. In the desktop app it can edit changes inline, review pull requests in a side panel, work across several repositories in one project, give each chat its own copy of a repository (a worktree) so parallel jobs do not collide, and run coding tasks in parallel in Codex Cloud. A goal mode keeps Codex working until it judges a goal complete or a token budget runs out, and Codex Remote lets you follow and steer a session from your phone.⁴³,²⁸,²⁹,⁴⁴,⁴⁵

Permissions and governance

You control what ChatGPT can reach, when it checks in, and which actions need your approval. In the desktop app you choose a permission mode:

• Ask for approval works inside the current workspace and asks before editing files outside it or using the internet.
• Approve for me keeps the same limits but sends approval requests to auto-review, a model that checks each action and asks you only about ones it finds risky.
• Full access removes the limits on files and internet use.
• Custom applies rules from a settings file.

In the cloud browser, permission to visit a website is separate from permission to take a consequential action on it, such as a purchase. In companies, administrators manage access, connected tools and permitted actions, and can see Work conversations and actions through the Compliance API.⁴⁶,⁴⁷,⁴⁸

Complex tasks use more of your plan than quick questions, and Codex draws on your ChatGPT plan's usage (reported).⁴⁹,²²

Briefing ChatGPT 2.0

OpenAI's advice for its GPT-6 models matches Anthropic's closely. State the goal, the relevant context, the constraints, the evidence required, the success criteria and the output format. State each instruction once. Ask for approval only after the worker has prepared a concrete result you can review. Start at medium reasoning effort and raise it only when the gain is worth the time and cost. To shorten a prompt that works, remove one group of instructions at a time and rerun the same tests.⁵⁰,²⁴

The same Four-Part Brief works on both: the brief you write for Claude is the brief you give ChatGPT Work. That is why this book teaches briefing once, not once per app.

Side by side

As verified on November 5, 2026.

Feature map

The open-source column describes OpenCode, which Part III covers beside Claude Code and Codex. "Not documented" means the feature was not found in the product's documentation, not that it is impossible.

What they now share

The two apps have converged. Both are now AI Worker platforms. Both have moved from answering questions to doing whole jobs: research, files, a browser, your computer, your connected apps, and finished documents, spreadsheets and presentations at the end. Both build presentations you can edit yourself and export to PowerPoint or Google Slides (ChatGPT's Slides is reported and still rolling out). Both draw charts and interactive visuals inside the conversation. Both let you save a way of working as a skill. Both keep a separate, specialist place for software. Both run work in the cloud so it continues with your laptop closed, and let you follow it from your phone. Both run work on a schedule. Both connect to outside tools through MCP, which is why a connector you build in Part III can serve both, and OpenCode too. And both vendors now give the same advice for briefing them: the outcome and how to check it, not the steps.

Where they differ

The two rows differ in the middle box, who decides that work begins. They end in the same place: a permission setting that decides what the worker may do once it starts. That is why this section begins with permissions.

How permission is set. This is the difference that matters most for an AI Worker. Claude sets one level, Auto or Manual, for a whole conversation, and the Code tab adds finer modes. ChatGPT sets a permission mode with a sandbox around the workspace, can route approvals to a reviewing model, and in the cloud browser separates permission to visit a site from permission to act on it. Before you delegate, read what each mode actually allows, not its name.

Who decides that work begins. In Claude, the conversation decides. In ChatGPT, you do, by choosing Work. This difference is narrowing (reported), so do not choose an app on it.²¹,²² Whichever app you use, the permission setting is where you control what happens once work begins.

What surrounds the coding agent. Claude's Code tab has grown into a workspace for running many sessions at once: nested sessions, a coordinating session and projects with cloud threads. Codex leans toward the software lifecycle: inline diffs, pull-request review, several repositories, worktrees and cloud runs.

What starts work without you. ChatGPT can start a task when an email, a Slack message or a pull request arrives. Claude's cloud tasks start on a schedule, a GitHub event or an API call.

What you can build on top. Claude's artifacts can call Claude for whoever uses them, at the user's cost. ChatGPT's Sites publish web apps that update from their data. Claude also lets you reshape the app itself with mods.

Where open source fits. OpenCode is a coding agent, not a desktop app for everyday delegated work. Its strength is choice: it runs models from many providers, including models on your own computer, which matters where cost, data location or vendor independence comes first. For a non-developer starting out, the two vendor apps remain the easier place to begin.

What the comparison cannot tell you

Neither vendor's documentation shows which app reasons better, makes fewer spreadsheet errors, finishes more web tasks, or keeps your data safer. An easier interface can reduce friction, but it cannot stand in for a correct result or a sound permission boundary. The only fair comparison is the one this book teaches you to run: the same job, the same inputs, the same done-when and the same limits on each, repeated, with the results checked against the files each run left behind, never against the worker's account of its own work.

Which model for which job

As verified on November 5, 2026.

Both apps let you pick a model, and both vendors give the same advice about it: the brief and the effort setting usually matter more than the model. Start with the model the vendor recommends for most work. Try a different model only when a job keeps failing its done-when at a higher effort setting, or when the job is routine, repeated often, or must be fast.⁵¹,²⁴,⁵²,⁵³

The models are specialists, not rungs on one ladder. A faster model can beat a more capable one at its own kind of job, so compare models on your own tasks before settling. Three things are also easy to mix up:

• The default: the model an app uses unless you choose another.
• What you can select: the other models the app lets you pick.
• What your plan includes: some models are available only on some plans, or use your limits faster.

The tables below say which is which.

By kind of job

Three cautions apply to any choice. A stronger model does not replace a human review: it makes better drafts, not accountable decisions. A model never does the final arithmetic in this book: calculations belong to tested code or formulas. And when you switch models, rerun your reusable briefs against their done-whens, because a brief tuned for one model may behave differently on another.

What matters more than any feature

Every feature on this page changes how work begins, where it runs and how long it can continue. None of them changes who is accountable for it. The book's first page, Claude and ChatGPT 2.0, sets out the four rules that matter more than any feature, and Part II teaches them: the Four-Part Brief, the Review Contract and the Authority Envelope.

Try it yourself. The practice lab on the first page has five short hands-on tasks, 10 to 20 minutes each, that you can do with a free account.

The Five Horizons comes next. It shows where this book is going, from one AI Worker to AI-native companies and countries.