Projects, Custom GPTs, and Gems: which reusable workspace to build

Overview

A reusable workspace is the container you build once — instructions plus a few knowledge files — and reuse across many chats, so you stop re-explaining context every morning. ChatGPT calls them Projects, OpenAI also has Custom GPTs, and Google has Gems.

Why now: ChatGPT Projects gained deep research in June 2025 and project-only memory that August; Gemini Gems landed in the Workspace side panel in July 2025; and in April 2026 OpenAI introduced workspace agents as “an evolution of GPTs”. Three roughly comparable containers — with the ground moving under one of them.

What you’ll be able to do: pick the right container for a recurring task, instead of defaulting to whichever brand you already pay for.

The content

The obvious read is that these are three names for the same thing. They mostly are — each is a saved set of instructions plus uploaded knowledge files. The differences that matter are about where the container lives and who else can use it.

ChatGPT Projects is a folder of related chats. Its distinctive move is isolation: when you turn on project-only memory, ChatGPT won’t pull your saved memories from outside the project into responses, and won’t carry project context back out into your other chats (OpenAI, Aug 2025). That makes Projects strong for keeping one client or matter walled off from everything else you do.

Custom GPTs are built to be published. The same instructions-plus-knowledge recipe, but packaged as a standalone assistant a whole team can run — with low fixed ceilings: around 20 knowledge files (OpenAI Help Center) and roughly 8,000 characters of instructions (community-reported; the official page states no limit). Two currents to know before you build one: Projects have been shareable on every plan since October 2025, so “I need to share it” alone no longer forces a GPT — its real edge is being a packaged, publishable assistant. And in April 2026 OpenAI introduced workspace agents as “an evolution of GPTs” for business plans; GPTs remain available, with conversion tooling promised, but that is where team assistants are heading.

Gemini Gems live where Google work already happens. Since the July 2025 rollout you invoke a Gem from the side panel of Docs, Sheets, Slides, Drive and Gmail. If your knowledge lives in Workspace, the Gem is the one that reaches it without a detour.

Here’s the limit worth internalising. A knowledge file is not the same as the model having read your file. All three retrieve relevant chunks rather than guaranteeing the whole document is in context (Redis, 2026) — and retrieval has a specific failure mode: documents that look relevant but don’t contain the answer measurably drag accuracy down, and a bloated workspace is a factory for exactly those near-misses (Cuconasu et al., 2024 — who found, curiously, that truly random text can even help; the enemy isn’t noise, it’s plausible-but-wrong context). The skill is curation. Three sharp files beat thirty.

Try it

Don’t start by building one. Start by deciding which kind you need, using a recurring task you actually have — a weekly report, a standing client, a document style you keep re-explaining. Paste this into any of the three assistants:

I have a recurring task I keep re-explaining to AI from scratch:
[describe the task in 2-3 sentences].

Help me decide which reusable workspace fits, by asking me:
1. Will other people run this themselves (a packaged assistant), or is it mainly mine?
2. Does it need to read files that live in Google Workspace?
3. Does it need to stay walled off from my other AI memory/chats?

Then recommend Projects, a Custom GPT, or a Gem, and tell me the
3-5 knowledge files worth uploading - and which to leave out.

Where this breaks: if the “context” is one long instruction you can paste at the top of a single chat, you don’t need a workspace at all. Building a container for a one-off is overhead, not leverage.

Additional reading

• Projects in ChatGPT — OpenAI Help Center — project-only memory and deep research within a project.
• Gems in the side panel of Google Workspace apps — Workspace Updates — the July 2025 rollout into Docs, Gmail and Drive.
• Creating and editing GPTs — OpenAI Help Center — the ~20 knowledge-file limit for Custom GPTs (living page, updated June 2026).
• Introducing workspace agents in ChatGPT — OpenAI (April 2026) — “an evolution of GPTs”: GPTs stay available, conversion tooling promised; the direction of travel for team assistants.
• How can I increase the ceiling for GPT Instructions beyond 8000 characters — OpenAI Developer Community — confirmation of the ~8,000-character instructions ceiling.
• The Power of Noise: Redefining Retrieval for RAG Systems — Cuconasu et al. — the nuance behind curation: related-but-irrelevant documents measurably degrade answers, while truly random ones can surprisingly help. The enemy is plausible-but-wrong context.
• RAG vs Large Context Window — Redis — why uploaded knowledge is retrieved as chunks, not wholesale read.

Editor’s note

The first thing I check in any of these containers is what leaks. In contract work, separation between matters is an obligation rather than a preference, so project-only memory matters more to me than any file limit. The second check is what you feed it. For example, giving a container the whole client folder feels thorough, but the folder is mostly drafts that look authoritative and aren’t — the four documents that actually define the matter will answer better. Pick the container that sits closest to your work; be deliberate about what it holds and what it remembers.