Know the Room

Claude is a large language model. Anthropic built it by having software read an enormous amount of the written world — books, articles, reports, conversations — until it got very good at one task: predicting which words come next. Run that prediction millions of times in a row and you get something that can draft a memo, summarize a report, or talk through a plan with you. Inside there is no database of facts and no little person. There is a machine that learned the patterns of how people write and reason.

Two things follow from how it was built, and they explain most of what you'll see in this program.

Claude knows what you give it. Your message, your files, and your Project's instructions are its entire world for that conversation. A vague brief gets you a generic draft. A specific brief gets you a usable one. The kickoff deck put it in six words: give it more, get more.

It can be wrong in a confident voice. Claude writes fluently whether it is right or not, because fluent writing is the thing it learned. When it has no good answer, it sometimes produces a plausible-sounding one anyway. People call this hallucination, and you will see it. The program's rules already cover the fix: you review everything before it ships, and the work goes out under your name.

Three facts about your data, because everyone asks. Your conversations are private by default. DC CAP's enterprise agreement means nothing you type is used to train the model. And Claude starts every new conversation from zero. It remembers nothing between chats unless you work inside a Project, where your team saves context on purpose.

The plain-English glossary covers the rest of the words people throw around: prompt, token, context window, hallucination.

The Claude Desktop App is one app with two modes. Chat is conversation: questions, drafts, thinking. Cowork is production: Claude creating real files on your computer. A toggle at the top of the app switches between them. Your DC CAP seat covers both, plus the same account at claude.ai in your browser.

Schematic: your sign-in screen will look close to this, with DC CAP's workspace name once you're in.

The sidebar is the building directory. New conversations start at the top. Everything you've ever asked lives below it, searchable. Projects (shared rooms you'll visit at Stop 4) have their own section.

The line that matters: your conversations are private by default. Teammates see your work only when you put it somewhere shared, like a Project. Your history is yours.

Talk to Claude the way you'd brief a capable colleague: context first, ask second. "I'm prepping a workshop for high school counselors next month — give me five icebreaker options that work in 10 minutes" beats "give me icebreakers" every time.

The second message is the real lesson. Everyone's first drafts need fixing, and asking for changes is the process working.

You can hand Claude a file (a draft, a PDF, a spreadsheet) and work on it together. Where outputs land depends on the mode: in Chat you download what Claude produces; in Cowork (Stop 5) finished files appear directly in your working folder.

A Project is a room your team shares. It holds two things: instructions (standing guidance Claude follows every time: voice, audience, format) and knowledge (files everyone in the room works from). Walk in, and Claude already knows the context. That's the point: nobody re-briefs from scratch. Those instructions are also where you tell Claude who it works for — the role it plays, the audience it writes to, the standards it holds. The more precisely you name that, the more it stays in character.

At the top of the app, a toggle switches Chat to Cowork. Cowork asks you to pick a working folder — the folder Claude can read from and save into. Ask for a Word document and an actual .docx lands in that folder. Decks, spreadsheets, and PDFs work the same way: real files, ready to open.

A lot of DC CAP work has a house way of being done — our emails, our board decks, the grant framing we've already approved. Skills teach Cowork those house ways, so when you ask for that kind of work the right one loads on its own and a small note names the Skill that's working. You never pick from a menu.

Nothing to set up today. Using and building Skills deliberately is a build-phase topic — the full walkthrough (what loads when, and how to brief a Skill like a colleague) arrives around Week 4. For now, just recognize the word, and know a note may appear naming a Skill when one activates.

Connectors let Cowork read from — and, with care, write to — tools DC CAP already runs on, like Monday.com, Google Drive, and Canva. They run on your existing access, so Claude only ever sees what your own login can see.

Nothing to set up today. Connectors — and the governance habit that goes with letting Claude change something in a shared tool — get their full walkthrough in the build weeks (3–4). For now, just know the word; which connectors are live for your seat is an Innovation Hub decision.

Three facts close the tour. Your history is saved and private by default. DC CAP's enterprise protections mean nothing you type trains anyone's model. And you can delete any conversation, which is the move you practice now, because deleting is also step one of incident response.