The short answer: pick one job your team repeats every week and set up a shared Claude Project or ChatGPT Project for it. Load the five to ten files a new starter would need, write one page of standing instructions and name one person who owns it, all in an afternoon. Don't build a new Custom GPT for this: OpenAI retires them on 11 December 2026.
Why does everyone's AI give different answers?
Because everyone briefs it differently, from memory, every morning. One account manager pastes in the brand guidelines and last month's best proposal. Another pastes in nothing and gets something generic, from the same tool.
Most companies respond by buying licences, which fixes access but not context. In a survey of 2,000 workers at companies with more than 500 staff, 49% said they'd used AI tools their employer hadn't approved, and 58% of those were on free versions (BlackFog, via CIO, January 2026; BlackFog sells security software). Approved or not, each person is running a private setup that only they understand.
A shared workspace moves the context out of people's clipboards and into one place that someone owns. That's the whole idea. The rest of this post is how to do it properly.
Licences fix access. A shared workspace fixes context.
Claude Projects, ChatGPT Projects or a Custom GPT?
Use a Claude Project or a shared ChatGPT Project, whichever your team already pays for. Don't start a new Custom GPT.
OpenAI told customers on 11 September that it will retire Custom GPTs on 11 December 2026 and replace them with plugins (Virtualization Review). A migrated GPT starts private, loses its chosen model and needs its sharing set up again (TechRepublic, October 2026). If your team already relies on one, migrate it, but don't build the next one.
Between the two kinds of project, these are the differences that matter for a team.
| Claude Project | Shared ChatGPT Project | |
|---|---|---|
| Plans that can share | Team and Enterprise | Every plan, Free to Enterprise |
| Files | Up to 30 MB each, no fixed count. On paid plans, Claude searches the files once they outgrow its context window | 25 per project on Plus, 40 on Business and Enterprise |
| Permissions | Can view, Can edit | Owner, Edit, Chat |
| Who sees your chats | Only you, unless you share one | Every member of the project |
| Memory | Each project keeps its own | Shared projects use project-only memory |
The row that catches people out is the highlighted one. In a shared ChatGPT Project, colleagues can read each other's project chats (OpenAI). That's often useful, because people learn from each other's prompts, but decide it's what you want before someone drafts a performance review in there.
On Google Workspace, Gemini Gems can now be shared like Drive files (Google, September 2025). Everything below applies there too.
What does the afternoon look like?
Four blocks of about an hour each, and the output is what we call Intersect's Workspace Brief: one job, the files behind it, the standing instructions and a named owner. If any of the four is missing, the workspace drifts.
| Block | What you do | Done when |
|---|---|---|
| 1:00Job and owner | Choose one piece of work three or more people do every week. Name the person who owns the workspace. | You can say the job in one sentence, and one name is on it |
| 1:30Files | Collect the current versions of what a new starter would need. Rename them so they're obvious. | Five to ten files, one version of each |
| 2:30Instructions | Write one page: the job, the reader, which file to trust for what, the always and never rules. | It produces a usable draft of one real task |
| 3:30Test and launch | Two colleagues rerun three tasks from last week. Compare the drafts with what was actually sent. Fix the instructions, then share the project. | Two of three drafts need only light edits, and the team knows where it lives |
Which job should the first workspace do?
One job, done every week, by at least three people. The weekly part gives you enough use to find the problems. The three people give you enough variation to show the value.
Good first jobs look like this:
- Proposals and pitch responses in your house structure and pricing.
- Client update emails and monthly reports in your tone.
- Tender and security questionnaire answers drawn from your policy library.
- New starter questions about how things work here.
Avoid "the marketing workspace" or "company knowledge". A workspace with a vague job gives vague answers, and nobody can tell whether it's working.
Which files should you load?
Load what you'd hand a new starter on their first day, and nothing you'd have to apologise for. In practice that's five kinds of file:
- The standard: two or three of your best finished examples of the output. Models copy an example more faithfully than a description.
- The facts: the current price list, service descriptions, product specs, who does what.
- The rules: the style guide, brand guide or policy that governs the work.
- The templates: the structure the output has to follow.
- The words: a short glossary of internal terms, product names and acronyms.
Leave out old versions sitting next to new ones, because the AI can't tell which is current. Leave out raw exports and whole folders, anything the whole group shouldn't see, and numbers that change daily.
Five to ten files is plenty to start with. Claude switches to searching the files once they outgrow its context window, which works (Anthropic), but more files isn't the same as better answers. Name each file so the instructions can point at it: "Price list - current.pdf" beats "final_v3.pdf".
What goes in the standing instructions?
One page that tells the AI what you'd tell a capable new starter: the job, the reader, which file to trust for what, and the rules. Here's an illustrative version for a proposals workspace. Copy it and change the specifics.
Job: You help the client services team at [Company] draft
proposals and pitch responses.
Reader: procurement leads and senior managers at mid-sized
companies. They skim. Lead with the outcome and the price.
Use the files like this:
- Prices: "Price list - current.pdf". Never quote a price
from anywhere else.
- Tone and structure: the three examples in "Winning
proposals". Match them.
- What we sell: "Service descriptions.docx". Don't offer
anything that isn't in it.
Always:
- Use British English.
- Keep the executive summary under 150 words.
- End with the next step and a date.
Never:
- Invent case studies, client names or results.
- Promise a timeline shorter than the service descriptions.
If the files don't answer something, say what's missing
and ask. Don't guess.
When you finish, list any assumptions you made.
Owner: [Name]. Last reviewed: [date].
Tell the owner when an answer is wrong.Paste this into the project's instructions field. It's the same in Claude and ChatGPT.
Three lines do most of the work. The file map stops the AI mixing an old price with a new one, and "if the files don't answer something, say what's missing" is the best single guard against confident, invented figures. And the owner's name at the bottom tells everyone who to go to when it's wrong.
Write it the way you'd brief a person. The same principles from our AI 101 on prompting apply, written once instead of every morning.
Who should own the workspace?
One named person who does the work the workspace supports. Not IT, and not a committee.
IT decides who can access what; the owner decides what the workspace knows. That's usually the most senior person who still does the job every week, because they'll spot a wrong answer first and care about it most. We run this blog the same way: one playbook that every draft is written and checked against, and one person who changes it.
The owner's job takes about 30 minutes a month:
- Swap in new versions of files and delete the old ones the same day.
- Keep a running list of bad answers from the team, and fix the file or instruction behind each one.
- Update the "Last reviewed" line whenever something changes.
- Close the workspace when the job changes. A stale workspace is worse than none, because people still trust it.
Give the owner edit rights and everyone else view or chat rights: Can edit and Can view in Claude, Owner and Chat in ChatGPT. If everyone can edit the instructions, they drift.
When doesn't a shared workspace work?
It stops working when it turns into the company attic. Fifty files and three versions of the price list produce answers that are confidently wrong, and the team stops trusting all of them. One job per workspace keeps it honest.
It's the wrong tool for live data, because project files are snapshots. Stock levels, pipeline numbers and today's rates belong in the system that holds them. ChatGPT Projects can search Google Drive and Slack as sources, which helps, but someone still has to know what it's reading.
It isn't private by default everywhere. Shared ChatGPT Project members see each other's chats, while Claude Project chats stay private unless shared. Neither is the place for HR cases or anything under an NDA that not every member has signed.
And it writes things; it doesn't do things. If the job is updating the CRM or sending invoices, you need an agent with access, permissions and guardrails, which is a different build.
What to do next
Pick the one job your team pasted the most context for last week. Block an afternoon, and name the owner before anyone opens the tool. The workspace without an owner is the one that's quietly wrong by the end of the quarter.
If your team uses Custom GPTs today, list them this week and decide which ones to migrate before 11 December.
Build your team's first workspace with us
We run hands-on sessions where a team leaves with its first shared workspace built, loaded, tested and owned. Book a call and bring the job you'd start with.
FAQ
Do we need a Team or Enterprise plan?
For Claude, yes: sharing a project needs a Team or Enterprise plan. ChatGPT lets you share projects on every plan, including Free, but Free and Plus have lower file and collaborator limits, and the work sits in personal accounts rather than a company workspace. For anything with client data, use a business plan your company controls.
What happens to the Custom GPTs we already use?
From 11 December 2026 they stop taking new conversations, or from 11 February 2027 for Enterprise workspaces with an approved deferral. Each one can be migrated to a plugin: its instructions become a skill and its files carry across, but custom actions don't. Old conversations stay readable.
How many workspaces should a team have?
Start with one. Add the second only when the first has an owner, a review date and people using it every week. After that, think one per recurring job, not one per person.