Start here
Welcome to Claude at Casper. This page teaches you the basics: the vocabulary, which tool and model to use, how-to guides, tips, and ready-made prompts for your team. Follow these five steps in order. Together they take about 35 minutes.
Using Claude safely
Do
- Sign in with your work email, not a personal account.
- Check facts, numbers, prices, and policy claims before anything leaves your hands.
- Keep a named person accountable for every customer-facing output.
- Tell Claude where the source is: attach it, link it, or connect the app.
- Share prompts that work with the rollout team so we can add them to your team's playbook on this page.
Don't
- Paste payment card numbers, passwords, or API keys anywhere.
- Share customer personal data unless the tool and use are approved.
- Put confidential Casper data into personal or unapproved AI tools.
- Publish or send AI-written content externally without a human review.
- Let AI make final calls on people decisions like hiring or performance.
Casper's official AI and data policy always comes first. If anything here is unclear, ask before you share data with Claude.
The words you'll hear
Plain definitions, each with a Casper example. Type to filter.
Chat, Cowork, Claude Code, and more
The model is the brain. Each surface gives it different hands: a conversation, your files and apps, a code repository, a browser, or a spreadsheet. In the current Claude app, chat and Cowork are one place. You don't switch modes, you just say what you need.
Chat vs. Cowork vs. Claude Code at a glance
| Chat | Cowork tasks | Claude Code | |
|---|---|---|---|
| Think of it as | A sharp colleague you talk to | A colleague who does the task for you, start to finish | An engineer working in your codebase |
| How to start | Type in any conversation | Describe the finished outcome in any conversation. On older app versions, pick Cowork in the message box first. | The Code tab in the desktop app, or run claude in a terminal |
| You give it | A question, a draft, attached files | A goal, files or folders, connected apps | A repo and a coding task |
| You get back | Answers, drafts, analysis in the chat | Finished files: docs, sheets, decks, reports, plus recurring scheduled tasks | Code changes, tests, commits, pull requests |
| Best for | Thinking, writing, quick analysis | Multi-step work across files and apps | Building and fixing software |
| Who at Casper | Everyone | Everyone, once comfortable with Chat | Engineering, data, e-com developers |
Pick the model for the job
The model menu sits next to the send button. Click the model name to switch models (choose More models to see the rest), set Effort, or toggle Thinking. Deeper models are slower. Faster models are fine for simple, repetitive work. If you're not sure, use Opus.
Fable handles your data differently
- Every request is safety-checked. Automated checks review everything Fable reads, including files, connector data, and web results. They look for a narrow set of topics: hacking techniques, risky biology, attempts to copy the model, and some AI-model development work. Everyday Casper work shouldn't trigger them. If a request is flagged, Claude re-runs it on Opus.
- Data is kept for at least 30 days. Fable prompts and outputs are then deleted automatically, unless they were flagged or must be kept for legal reasons. Zero data retention isn't available for Fable.
- Access is restricted. Anthropic staff can't read stored conversations by default. Review happens only through a controlled path, and every access is logged.
- No training by default. On business plans, Anthropic doesn't train models on your inputs or outputs. Thumbs up/down feedback is the exception.
Our guidance: until IT and Legal confirm otherwise, use Fable for hard reasoning on non-sensitive work. Keep customer personal data and confidential financials on Opus or Sonnet.
Source: Anthropic help center, "Covered Models" and "Data retention practices for Covered Models" (October 2026).
What are you trying to do?
Apps Claude can work with
A connector lets Claude read from, and sometimes act in, an app you already use. It always uses your own login and permissions. If you can't open something yourself, Claude can't either. These connectors are approved for Casper's Claude organization.
How to turn on a connectorAbout 2 minutes, once per app
- In the left sidebar, open Customize → Connectors, then click + next to Connectors. In a conversation you can also click + (or type /), hover over Connectors, and choose Manage connectors.
- Pick the app below and click Connect.
- Sign in with your Casper account and approve access. Everyone signs in for themselves, even when the app is approved for the company.
- In a conversation, click +, hover over Connectors, and switch on the apps you want Claude to use there.
- Ask for what you need. For example: "Search the #cx-escalations Slack channel for this week's top three issues."
Don't see an app you need? Send the request to the rollout team with the use case. A Claude organization Owner adds new connectors for the whole company in Organization settings → Connectors.
Connectors to evaluate next
Ask each team which systems they live in. Common candidates for a DTC sleep brand: the customer support help desk, email & SMS marketing, BI dashboards, design files, ERP/finance, and HRIS. Search the Connectors directory in Settings and send the strongest cases to IT.
Step by step
Start at the top if you're new. Each guide stands on its own.
Real examples for each Casper team
Copy a prompt, replace the [brackets], and go. Each example lists the tool, model, and connectors it uses.
Habits that get better results
Make your usage limit go further
Every message counts against your usage limit. Long conversations, connectors, web search, higher Effort, and deeper models all use more tokens. A few habits make a big difference.
Do
- Start a new conversation when the topic changes or a chat gets long. Long chats use more of your limit on every turn.
- Use Projects for documents you reuse. Projects pull in only the parts that are relevant to your question.
- Keep Project instructions short and remove Project files you no longer use.
- Switch off connectors and web search you don't need for that conversation. Tools and connectors use a lot of tokens.
- Match model and Effort to the task. Use Sonnet or Haiku with lower Effort for routine rewrites and summaries.
- Say how long the answer should be, for example "5 bullets" or "under 100 words."
- Put it all in one message: context, task, format, and every question you have.
Don't
- Keep one giant chat going for weeks.
- Paste the same file or long text again and again. Attach it once, or keep it in a Project.
- Leave every connector switched on "just in case."
- Use Fable or Max Effort for quick rewrites or simple questions.
- Ask for a full rewrite when you only need one fix. Say "change only the second paragraph."
- Send a string of small corrections. Edit your original message and resend it instead.
Fewer, sharper skills work best
We're still deciding how to open up and govern skills and plugins at Casper. Check with the rollout team before you add a plugin or share a skill.
Claude sees the name and one-line description of every skill you've turned on, and picks one based on that description. It reads the full instructions only when a task matches. There's no official limit on how many skills you can have. As a rule of thumb, start with 3–5 per team and add more only when a task keeps repeating.
Do
- Turn on only the skills you use. Turn off the rest.
- Give each skill one clear job and a specific description that says what it does and when to use it. For example: "Writes CX macros in Casper's voice. Use for any customer reply template."
- Keep the main instructions short (under 500 lines). Put long reference material in separate files. Claude reads those only when it needs them.
- Test a new skill on 3 real tasks before you share it with your team.
- Read a shared skill before you turn it on. Its instructions and scripts run inside your conversation.
Don't
- Install overlapping skills that do the same job, like three different brand-voice skills. Claude may pick the wrong one.
- Write vague descriptions like "Helps with documents."
- Put every policy and guideline into one giant skill.
- Turn on skills from unknown sources without reading them.
Based on Anthropic's help center ("Understanding usage and length limits") and Skills documentation, October 2026. The 3–5 skills per team is our starting guidance, not an Anthropic limit.
Where are you today?
10 questions to check what you've learned. Answer them all, then check your answers to see your score and why each answer is right. Nothing is saved or sent anywhere.