/
Blog
Back to Blog
How-To

ChatGPT Projects for a Team: Shared Instructions in 90 Seconds

Abo-Elmakarem ShohoudNovember 3, 20264 min read

Put ten people on a team in front of ChatGPT and you get ten different replies to the same client email. Different tone, different promises, and sometimes a discount nobody approved. A ChatGPT Project fixes most of that: one set of shared instructions and reference files that every chat inside the project starts from. Here is a 90-second setup, tested on one real client question.

The problem, on one question

In the video, the same client email goes to ChatGPT twice:

Hi, we've had to close the office tomorrow. Can you move the clean to next week? And as it's such short notice, could you take 10% off this month's invoice?

In a plain chat, ChatGPT wrote a polite reply. It agreed to reschedule but asked the client to pick a day, said nothing about the late-notice fee, and refused any discount because "resources were already allocated". Nothing in it was rude, but almost all of it was wrong for this company.

Inside the project, the reply named clause 4.2 and the £35 rescheduling fee (and that it can be waived once a year), offered two real visit windows from clause 4.3, and capped the goodwill credit at 5% under clause 6.2, saying the full 10% request would be checked with a team lead. Same model, same question. The difference is the project.

How to set it up

  1. Create a project. In the sidebar, open Projects and add a new one. Give it a name the whole team will recognise, such as "Customer Replies".
  2. Add shared instructions. Open the project's menu, then Project settings, and paste the instructions into the Instructions box. Save.
  3. Add reference files. On the project page, open Sources, then Add sources, and upload the documents the answers must follow.
  4. Share it with your team. Use Share on the project page and invite people. Every chat they start in the project uses the same instructions and files.

The instructions used in the video

You draft replies to client emails for Brightwell Facilities, an office cleaning company in Manchester.
Use warm, plain British English. Keep replies under 120 words.
Check the Service Terms in this project before promising anything, and name the clause you used.
Never offer a discount above 5%. For more, say you'll check with your team lead.
Offer two concrete visit options from the terms.
Sign off as: Best wishes, [Name], Brightwell Customer Care

Each line has a job:

  • Who it is for and the tone keeps every reply sounding like one company, not ten people.
  • "Check the Service Terms ... and name the clause" is the line that matters most. It forces the model to use your files and makes every promise checkable.
  • The discount rule turns a judgement call into a policy. Anything above the limit goes to a person.
  • "Two concrete options" stops the vague "let us know what works" replies that start a second email thread.
  • The sign-off is small, but it is what clients notice first when it varies.

The reference files

The project in the video has two files, both fictional and made for the demo:

  • Service terms with numbered clauses: visit windows, prices, the 48-hour rescheduling rule, the £35 fee, and the 5% goodwill limit.
  • Approved reply examples: two real-looking replies that show the structure and tone you want, with the clause named in the text.

Two or three good examples do more than a long style guide. Keep the files short, current and numbered, because the model will quote them.

Check before it goes out

A project gives everyone the same starting point. It doesn't replace judgement:

  1. Read every reply before you send it. Check the clause it names actually says that.
  2. Keep the terms file current. When a price or rule changes, replace the file in the project the same day, or the model will keep quoting the old one.
  3. Only upload what your company allows in its approved AI tool. Client names, contracts and personal data need a plan your company has cleared for them.
  4. Review a sample each week. Read five replies from the team and tighten the instructions where they drift.

Why this belongs in AI training

This is one of the first things I set up with teams in hands-on sessions, because it fixes the problem managers notice first: everyone using ChatGPT differently. Once a team has one shared project per recurring task (client replies, proposals, reports), the quality of the output stops depending on who wrote the prompt.

If you want your team trained this way on your own documents, see remote AI training for teams, or estimate a budget with the AI training cost calculator.

Share this post