TaskbyTask.Practical AI guides for work
ChatGPT or Claude · Practical walkthrough

Set up your first AI workspace

It's your first day with ChatGPT or Claude at work, and the first question isn't what to ask it. It's which account you're actually signed in to, and what happens to what you type.

Task by Task13 min read + practice
An organised workspace tray, project folder, key tag and settings dial illustrate setting up an AI workspace.

What you'll make

one checked practice response in the account you mean to use, and a short setup note recording what you checked.

What you'll need

your existing ChatGPT or Claude account, your organisation’s AI rules, and about fifteen minutes. If you have a workplace invitation, use its sign-in route. Don’t start another subscription just because the first login is elusive.

Example result.

Customer sentence: Your boiler service visit is booked for Thursday between 13:00 and 15:00.
Confirm before sending: the calendar date; the engineer's name; how the engineer will get access.

1. Check which account you’re in

Open the account menu and note three things:

  • the email address you’re signed in with;
  • whether a workspace or organisation name appears;
  • the plan: Free, Plus or Pro, Team or Business, or Enterprise.

A project called “Company work” in a personal account is still in that personal account. Its name proves nothing. If you’re unsure which account to use, ask your administrator, and don’t merge or move workspaces for this exercise.

Why this works: the account, not the chat, decides who controls your data and which settings apply.

What you should see: an email address, a workspace name (or none) and a plan name you can write down.

2. Review the data controls

On a personal account, check whether your chats can be used to improve the models:

  • ChatGPT: Settings › Data controls › Improve the model for everyone.
  • Claude: Settings › Privacy › Help Improve our AI models.

Turning either off stops new chats being used for training. It doesn’t delete your chats or make every work file suitable to upload. On business accounts (ChatGPT Business or Enterprise, Claude for Work), the providers say content isn’t used for training by default, and your administrator manages the controls.

Why this works: training, retention and sharing are separate questions, and this toggle only answers the first.

What you should see: the setting in the state you’ve chosen, or a note that your administrator manages it.

3. Start a new chat and try this

Fictional practice note:
The boiler service visit is booked for Thursday, 13:00–15:00. The engineer's name and the access arrangements have not been confirmed.

Turn this into one short customer-facing sentence. List anything we must confirm before sending it. Use only the note. Do not invent missing details.

You don’t need a connected drive, calendar or API key.

Why this works: the note has one fact to keep and two gaps to flag, so you can check the answer in seconds.

What you should see: one sentence and a short “confirm before sending” list.

4. Check the answer

Check that the time survives exactly and that nothing missing has been filled in: no engineer’s name, no access instructions, no calendar date the note didn’t give.

A correct answer doesn’t prove your privacy settings are right. Check those in the interface, as in step 2.

Why this works: the first habit worth building is reading an answer against its source, even a three-line one.

What you should see: the example result at the top of this page, or something close to it.

5. Write a setup note before you use real work

Keep this outside the chat, and fill in “Data controls reviewed on” today:

Tool and account:
Workspace checked:
Information I may use:
Information I must leave out:
Sign-in and recovery checked:
Data controls reviewed on:
Practice project and sharing checked:
Person to ask if unsure:

Keep passwords, recovery codes and API keys out of chats. Add a connection to email, storage or calendars only when a specific task needs it, after checking what it can read and change.

Why this works: next month, the note answers “which account was I meant to use?” faster than memory does.

What you should see: a filled-in note saved somewhere you’ll find it.

Check the result

  • The account email and workspace match the ones you meant to use.
  • The data-control setting is in the state you chose.
  • The practice answer kept 13:00–15:00 and invented nothing.
  • The setup note is filled in and dated.

The usual failure: the answer fills a gap to make the sentence read better, such as “Our engineer, Sam, will arrive on Thursday 15 October” or “please leave the side gate open”. Spot it by checking every name, date and instruction against the note.

Your sentence includes [invented detail], which is not in the note. Remove it. Keep only what the note states, and add the missing item to the "confirm before sending" list.

[invented detail] is the part you couldn’t find in the note. Once you know which account and information you can use, try your first useful prompt.

Worked example: the helpful engineer

Fictional example. The note from step 3 produced:

“Good news! Our engineer, Sam, will visit on Thursday between 1 and 3pm. Please make sure someone is home to let them in.”

  • The failure. “Sam” isn’t in the note; the model invented a name to sound friendly. “Someone is home” turns an unconfirmed access arrangement into an instruction. The time survived, reformatted.
  • The repair. The repair prompt, with “the engineer’s name, Sam, and the access instruction”, gave the example result at the top of this page.
  • Why it matters. On a fictional note it’s harmless. On a real customer message, Sam doesn’t exist and the customer waits in for nothing.

Task kit for this job: Download (ZIP) · Reusable instructions

Go deeper

Optional training: Protect your sign-in and recovery

For ChatGPT, open Settings › Security and login, then under Multi-factor authentication (MFA) choose a method, such as an authenticator app or a passkey, and follow the steps. Once on, it applies across OpenAI services. Set up a recovery option you can reach.

For Claude, protect the email or Google account you sign in with, including its own multifactor authentication. If your organisation uses single sign-on, follow its requirements.

Don’t share a sign-in between colleagues. Use individual invitations where your plan supports them.

Optional training: Know what the data controls don't cover

A training setting is one part of data handling. Retention, sharing, administrator access and any connected service are separate questions, and a toggle can’t approve a client file on your behalf.

Temporary chats have their own rules too. ChatGPT’s Temporary Chat and Claude’s Incognito chats (the ghost icon at the top right of a new chat) aren’t saved to your history or memory. That doesn’t mean they’re deleted instantly: Claude keeps incognito chats for 30 days by default. Check each provider’s explanation before relying on either.

Optional training: Create a private practice project

A normal chat is enough to start. A project helps when you want the same instructions across repeated tasks. In ChatGPT, choose New project in the sidebar and add instructions in the project’s settings. In Claude, open Projects › New Project, keep it private and use Set project instructions. Leave the project unshared; check its sharing controls rather than trusting its name.

This is a practice space using fictional business examples. Help me make clear drafts in plain English. Use the source material I provide for factual claims. Keep names, numbers, dates and commitments unchanged. If information is missing or contradictory, point that out. Separate a suggested improvement from a stated fact. Ask before making an assumption that changes the meaning.

These instructions guide the output; they’re not a security control. For keeping projects apart, see Use ChatGPT Projects to keep one piece of work out of your other chats.

Sources

Optional analytics

With your permission, Google Analytics uses cookies to measure visits and which guides people read. It stays off until you accept. Use Analytics choices in the footer to change your choice. Rejecting after accepting refreshes this page. Privacy details.