Claude, screen by screen.

He opens with a finished research report and points at the Code tab. The name sounds like a black terminal for engineers. Today it lives inside the Claude desktop app and looks like the normal chat box.

In the Code tab he opens Artifacts, starts a new doc and types one line: do deep research on cryptocurrency, run five parallel sub-agents. Claude writes itself a to-do list first, then a doc opens and fills up while the research runs. A Background tasks panel shows five helpers, how many are done and which model each uses.

Three real builds from Reddit and X. One man replaced the ad-filled Samsung TV home screen with a clock and only his apps. Another wired a USB lamp to Claude hooks so it lights up when Claude needs him. A third rebuilt a slow Plex app for his LG TV. None of them are engineers. They described what they wanted.

Back in the normal chat. He shows the model picker and effort level in the box, the permission picker next to it, dictation, voice mode, and Search, which looks through chats, artifacts, sessions, projects and scheduled tasks at once.

He opens a project used for his community resources. Every chat inside shares the same instructions, files and folder. He then makes a new one called test.

The Artifacts page shows every doc, deck, design and dashboard Claude made, as tiles you can share by link. It works the same in chat and in the Code tab.

A morning brief appears at 8 AM with the day's calendar and important emails. Nobody typed anything. To set it up he clicks Scheduled → New task → Create with Claude and types one sentence. Claude asks him to connect Gmail and Calendar and asks two questions.

He opens Discover with 57 categories of ready-made skills (Engineering alone has over 1,200). Under Customize he switches on built-in skills. Then Add → upload, create with Claude, or record your screen. On a Mac Claude watches the recording and turns it into a skill.

He shows a 5-slide deck on Claude's top slash commands, all in one style, made in the Design tab. If Design is not in the sidebar: More → Edit sidebar → drag it up → Done. Inside he builds a design system: colours, type, spacing, corner radius, components, accessibility and icons. Then asks for the deck on Opus 5.5 at medium effort. Every slide stays editable.

Customize → Connectors → Gmail → Add. A browser opens, he logs in and comes straight back connected. Then plugins: he copies the GitHub link of a design plugin, goes to Customize → Plugins → Create with Claude, pastes the link, picks Install and chooses both the Claude app and Claude Code.

Same chat box, but Claude works inside your files. He picks where it runs: Local (your computer and files) or Cloud (keeps going when the app is closed). Then he picks a folder. The folder is Claude's desk. Docs, slides and design work on the free plan. The Code tab needs Pro or above.

He opens the permission menu. Manual asks before every edit, command or web visit. Accept edits changes files in your folder without asking. Plan looks around, writes a plan and changes nothing until you approve. Auto skips routine questions while a second model checks each action. Example: clean up the Downloads folder. In Plan mode it first lists 40 files, invoices to one folder, photos to another, old installers deleted. Bypass in settings asks nothing. Anthropic says use it only on a spare test machine.

In the Code tab he opens Artifacts, picks the design system made earlier and types: create a presentation on Claude Cowork vs Claude Code. Environment is set to This computer, pointed at one new folder in Downloads. He skips Fast mode because it costs about double from extra usage credits.

He types /remote-control on the laptop, opens the Claude app on his phone, and the same session is there, still working. He sends the next instruction from a taxi in Bengaluru traffic.

The Context panel shows how full Claude's whiteboard is (up to a million tokens). When it fills, Claude squeezes older parts into a summary by itself. He also shows the connectors panel, a terminal and a git view (not needed), the built-in browser (separate from your Chrome, so none of your logins) and the Files pane with every file Claude made.

While the deck builds he opens a side chat with Cmd + ; on Mac (Ctrl + ; on Windows) and asks what it is covering. Answer in seconds: slides 1 to 4 done, 5 to 9 planned. The main job never stops and nothing from the side chat goes back into it.

Mods came out at the start of October. He opens a new session and types: create a mod that shows a progress bar and the estimated cost while a task runs, trigger it with /agent-info. Claude writes the mod and reloads the app. The bar now appears when a prompt ends with /agent-info and disappears when the job is done.

Here a project starts from a folder. Every chat inside shares that folder and its files. He makes a section called Crypto, moves an old chat into it and adds a new one asking what Bitcoin is. Scheduled tasks and Customize work the same as in chat.

He types a slash and his own YouTube editing skill pops up. Then seven commands, least used first: /advisor (a stronger model gives a second opinion at key moments), /clear (fresh session, empty memory), /usage or /cost (what the session cost), /doctor (checks your setup and fixes it), /deep-research (research job on a topic), /btw (side chat) and /goal.

Six sessions on one screen, each on one part of the same YouTube video: thumbnail, title, outline, visuals, editing. He typed one prompt in the main session asking it to start a thread for each part. The main session acts like a producer. Each nested session has its own memory, so nothing gets mixed up. Drag sessions across, or ⋯ → Open in split view, to watch all six.

He ends with three steps to stop using Claude like a search box.