You wrote a good set of instructions once, pasted them into a chat, and then had to paste them again the next day. All three big assistants now have a saved-context feature to fix that, and the names are different enough that comparison posts get confusing. Here is how Claude Projects, ChatGPT Projects and Gemini Gems differ on the points that matter, using vendor pages I read on 2026-10-09, and what to do about the parts I could not read.
Short version: pick by where your work already lives and who needs to share it, not by feature count. And keep a copy of your instructions outside any one tool, because one of these three is being renamed.
How were these facts checked?
| Vendor | Page | Status on 2026-10-09 |
|---|---|---|
| Anthropic | "What are projects?" on support.claude.com | Read in full by curl |
| "Use Gems in Gemini Apps" on support.google.com | Read in full by curl | |
| OpenAI | "Projects in ChatGPT" on help.openai.com | HTTP 403 by WebFetch and curl. Not read. |
For ChatGPT I ran a web search and got secondary sources (third-party blogs). They disagreed with each other in places, so I treat ChatGPT numbers below as unverified and I do not rely on them. If a number matters to you, open OpenAI's help article yourself.
What does each feature actually give you?
| Claude Projects | ChatGPT Projects | Gemini Gems | |
|---|---|---|---|
| Unit | Workspace with its own chats and knowledge | Workspace grouping chats, files and instructions | Saved custom assistant |
| Standing instructions | Yes, project instructions | Yes (see the setup guide) | Yes, name plus instructions |
| Reference files | Upload to project knowledge | Yes; per-plan file caps unverified | Upload, add from Drive, or add a NotebookLM notebook |
| Drive files stay current | Not stated on the page I read | Not verified | Yes: Gemini uses the most recent version of a Drive file |
| Large knowledge sets | Paid plans: RAG mode expands capacity up to 10x when knowledge nears limits | Not verified | Limits not on the page I read |
| Sharing | Team and Enterprise: can view or can edit, by person, by list or org-wide | Not verified | Can share a Gem or a conversation with it |
| Free plan | Yes, up to five projects | Not verified | Not stated on the page |
| Future | Anthropic describes a beta new version of projects, rolling out in stages | Not verified | Gems become skills on personal accounts from November 2026 |
All Claude and Gemini entries come from the vendor pages on 2026-10-09 as I read them. "Not stated" means the page I read did not say, not that the answer is no.
Claude Projects
Anthropic's article says projects are self-contained workspaces with their own chat histories and knowledge bases, available to all users, with free accounts capped at five projects. On paid plans, RAG kicks in automatically as knowledge approaches context limits, expanding capacity by up to 10x according to the page. Team and Enterprise plans get sharing with two permission levels, and owners can turn sharing off.
One more thing from that page: a new version of projects is in beta, starting with Claude Code, where a project is one conversation that Claude splits into parallel threads. Existing projects keep working as they do today. That is a rollout, not a change you need to act on. For how to structure a project well, see the existing Claude Projects guide.
ChatGPT Projects
I could not read OpenAI's page, so I will not quote limits. What I can offer is a pointer: my existing ChatGPT Projects, Memory and Custom Instructions setup guide covers how the layers interact. Secondary articles quote file caps per project that differ by plan and, in one case, flag that OpenAI's own pages disagree with each other. That is a good reason to check the live page before you design around a number.
Gemini Gems
Google's page says you create a Gem from the Gems section in the Gemini web app, add a name and instructions, optionally add knowledge files, and can have Gemini rewrite your instructions. A Gem created in the web app also appears in the mobile app and the Gemini side panel in Workspace. Gems cannot be used with Gemini Live at present.
The headline is a banner on that page: starting in November 2026, Gems go away on personal Google accounts and transition automatically to skills. Google's separate transition article (also read 2026-10-09) says work and school accounts follow in 2027, that skills are reusable instructions you can call with a slash command or let Gemini apply automatically, and that you can import SKILL.md files. For the practical consequences and a build guide, read the Gemini Gems guide and Google's own transition article.
Which one should I choose?
Decide with these questions, in order.
- Does a team need to edit the same context? Claude's Team and Enterprise sharing has explicit view and edit levels. Gemini lets you share a Gem. For ChatGPT, check your plan's sharing page. If the answer is "wiki-like and many authors", compare against a notes tool too: Notion AI vs Claude Projects.
- Do your source documents live in Google Drive? Gems can reference Drive files and pick up edits, which saves re-uploading. That is a real advantage if your reference material changes weekly.
- Is your knowledge set big? Claude documents RAG expansion on paid plans. For the others, I could not find a documented figure on the pages I read.
- Do you want to avoid a rebuild next quarter? Gems are the only one of the three with a dated transition notice. Do not sink hours into elaborate Gems before reading it.
- Is this free-tier use? Claude: five projects. Others: check the live page.
If none of those decide it, pick the assistant you already use daily. Switching cost is mostly retyping, which the next section removes.
How do I make my setup portable?
Keep one plain text file per use case. This is an untested template, but it is just text. Paste the sections into whichever field the tool offers.
NAME: Client proposals
ROLE: You draft proposals for a small design studio.
AUDIENCE: Non-technical founders.
STYLE: Plain English, short paragraphs, no jargon, UK spelling.
ALWAYS: Ask for budget and deadline if missing. List assumptions at the end.
NEVER: Invent past clients or prices. Write "NOT PROVIDED" instead.
OUTPUT: Headings: Summary, Scope, Timeline, Price, Next step.
REFERENCE FILES: pricing-sheet.pdf, past-proposal-1.docx (names only; re-upload per tool)
Then run the same three test prompts in each tool and compare the answers side by side:
- A normal request that should follow every rule.
- A request missing the budget (does it ask or invent?).
- A request that tempts a made-up client reference (does it write NOT PROVIDED?).
Keep the winner. This costs ten minutes and tells you more than any feature table, including this one.
What should I do this week?
- If you use Gems, read Google's transition article and copy each Gem's instructions into a text file now.
- If you are on a free Claude account, count your projects against the five-project cap.
- If you use ChatGPT Projects, open the live help article and note the file cap for your plan with the date.
- Write the portable brief for your top use case, and test it in two tools.
What I could not verify
- Every ChatGPT Projects number: file caps, project counts, sharing rules and plan differences (OpenAI pages returned 403).
- Gemini file-size or count limits for Gems (the page linked to a separate limits article I did not read).
- Whether Claude shows Drive-style live file sync: the page did not say.
- Exactly when work and school accounts move (Google's transition article says 2027, without a date).
All other statements come from the Anthropic and Google pages on 2026-10-09. Vendor pages change often, so recheck before you commit to a plan.
