So here's a scene I've watched play out in every company I've worked at, from dot-com startups to open source foundations. Someone asks in Slack, "Hey, what's our process for X?" The person who knows is in a meeting. Three people give three slightly different answers. Two days later, someone else asks the same question.

Institutional memory leaks. Always has. We've tried intranets, Confluence wikis, Notion databases, Google Docs, Slack threads. The stuff gets written down. Nobody reads it. The knowledge still lives in people's heads.

Claude Projects and ChatGPT Projects are the first AI features I've seen that actually solve this — not because the tech is magic, but because they change the ask. Instead of "go read our 47-page onboarding doc," the ask becomes "ask Claude" or "ask ChatGPT." And the AI reads the doc for them.

Both vendors shipped the same idea within months of each other. Claude Projects launched in mid-2024. ChatGPT Projects followed in December 2024 and rolled out to free users in April 2025. They work almost identically. Which one you pick matters less than whether you actually set one up.

Claude Projects and ChatGPT Projects turn the stuff nobody documents into something your whole team can query.

A Claude Project is a workspace that holds up to 30MB per file with no limit on the number of files, plus a custom instructions field that shapes every conversation. On the Team plan, a Project can be shared with everyone on your team.

A ChatGPT Project does the same thing — a folder for chats, files, and project-specific instructions, with memory scoped to that Project so context doesn't leak across your other work. File counts scale by plan: 5 on Free, 25 on Plus, 40 on Team/Pro/Enterprise.

Anyone inside a Project gets the same assistant — same knowledge, same instructions, same tone. That's the whole pitch. The execution is where most teams fumble.

In Claude:

Step 1: Log into claude.ai with a Claude Pro ($20/mo) or Claude Team ($30/user/mo monthly, $25/user/mo billed annually) account. Projects is not available on the Free plan.

Step 2: Click "Projects" in the left sidebar, then "Create Project."

Step 3: Name it something your team will actually find. "Sales Operations" beats "Project 1."

In ChatGPT:

Step 1: Log into chatgpt.com. Projects works across Free, Plus, Team, Pro, and Enterprise — OpenAI opened it to every tier in April 2025.

Step 2: Click the "+" next to "Projects" in the left sidebar.

Step 3: Name it, pick a color, and optionally drop in instructions on the same screen.

Time: 2 minutes on either.

The instinct is to dump every document you own into the Project. Don't. Claude works better when the knowledge is curated than when it's comprehensive.

For a sales operations Project, upload:

Skip the 200-slide product deck from 2023. Skip the internal Slack export nobody needs. Skip anything that's outdated — Claude will cite stale facts if you hand it stale sources.

I learned that the hard way. First Project I built had forty docs dumped in and Claude kept pulling the wrong pricing sheet. Trimmed it to five current files and the answers snapped into focus. Curation beats volume every time.

Step 1: Click "Add Content" in the Project sidebar.

Step 2: Upload files one at a time or drag a folder. PDFs, Word docs, Markdown, plain text, and code files work in both tools. Claude caps each file at 30MB. ChatGPT allows larger individual files (up to 512MB) but cuts the file count by plan as noted above.

Step 3: After upload, click into each file and confirm the assistant parsed it correctly. Both Claude and ChatGPT stumble on oddly formatted PDFs — multi-column layouts and scanned images are the usual culprits.

Time: 10 minutes for a well-curated knowledge pack.

Limitations: Claude Projects uses a 200K token context window (roughly a 500-page book). Past that, Claude switches to retrieval mode and pulls the most relevant chunks. ChatGPT Projects rely on retrieval across whatever files you've added, with a tighter hard cap on file count by tier. Both are good-but-not-perfect. Niche questions on documents the assistant didn't retrieve will get weaker answers.

This is the step most people skip. It's the one that matters most.

Custom instructions tell the AI who it is, who it's talking to, and what it should do when asked a question. Claude calls this field "Custom Instructions." ChatGPT calls it "Instructions" inside the Project (separate from your global ChatGPT instructions). Same function. Treat it like the onboarding packet you would give a new hire on day one.

Working template:

Step 1: In a Claude Project, click "Custom Instructions." In a ChatGPT Project, open the Project and click "Instructions" (or the gear icon).

Step 2: Paste the template. Fill in the brackets. Swap "Claude" for "ChatGPT" on the first line if you're using OpenAI.

Step 3: While you're in ChatGPT settings, set the Project to project-only memory if you're working on anything client-sensitive — it keeps the Project's context from leaking into the rest of your ChatGPT history.

Step 4: Test it. Ask a question you already know the answer to. Check that the model cites the right source.

Time: 10 minutes.

Claude: On Team or Enterprise plans, click "Share" in the Project and add teammates by email. They see the same files, custom instructions, and Claude behavior — no setup on their end. On Pro, Projects are private.

ChatGPT: On Team, Business, Enterprise, or Edu plans, you can share a Project with the whole workspace or pick specific people. On Plus and Free, Projects stay private to you.

Either way, a true team brain requires a paid team tier. Personal plans give you a private Project, which is still useful — but doesn't solve the tribal-knowledge problem you started with.

Be honest with your team about the limits. These apply to both Claude and ChatGPT:

If your team already lives in one of them, use that. If you're starting fresh: Claude Projects handles long documents better because of the 200K context window and the 30MB per-file cap. ChatGPT Projects are friendlier for mixed-media work and the free tier lets you prototype before you pay. Most teams I work with end up running both — Claude for deep document work, ChatGPT for quick tasks and images.

Pick the one question your team asks you most often this week. Build a Project around that single question — in whichever tool you already pay for. Five files, tight custom instructions, share with three people. If it answers the question 80% as well as you would, you have bought back an hour a week. Expand from there.

Don't try to build the full company knowledge base on day one. That's the intranet trap — ambitious, unused.

I appreciate your support.