How to Use Claude Projects: Upload Files and Instructions for Reuse

"I think I uploaded this same file last week."

Have you ever opened a new chat and found yourself digging up the same file, then explaining the same background all over again?

A project bundles your files, your instructions, and your conversations into one room. Create the room, upload your files, write your instructions, and every conversation you start there begins with the same material and the same rules already in place. No more re-uploading files or re-explaining context each time. Claude does this at claude.ai/projects, ChatGPT has a feature with the same name, and Gemini calls it Gems. If the previous lesson had you receiving output in an artifact window, today is about pinning the raw material that produces that output to a fixed room instead. If you're here specifically for how to write project instructions, this lesson alone covers it. Menu names and locations shift often, so if you're reading this much later, cross-check with each company's current documentation.

Project / Gem

A project is a workspace that holds your uploaded files, your pre-written instructions, and every conversation you've had inside it, all bundled together. Gemini calls the equivalent feature a Gem. Context that used to vanish the moment a single conversation ended now persists at the room level instead.

A project merges file attachments and custom instructions into one room

Let's start with the good news: there's almost nothing new to learn here. The previous track already had you doing both halves of this. You uploaded a file into a conversation to get something done. You also pinned your background and your answer rules into custom instructions. A project just merges those two things into one place.

The difference is scope. A file attachment lives and dies with that one conversation. Custom instructions sit at the opposite extreme: they apply to every conversation you have, which is too broad for a rule that only matters for one specific job. A project sits in the middle. It locks material and rules meant for one job into a single room.

Two things trip up beginners more than anything else here. Both are spelled out in the official documentation, but neither is obvious just from looking at the screen.

First: Claude never reads a project's name and description. Claude's official documentation states plainly that a project has a name and a description attached, but Claude doesn't access their content. Don't type "This project is for Client A's monthly report" into the description field and assume the AI picked that up. If you want Claude to actually read something, it has to go into project knowledge or into the instructions.

Second: conversations inside the same room can't see each other in Claude. Claude's documentation is explicit that conversations within a project don't share context unless something has been added to project knowledge. Don't expect conversation 2 to remember the report format you settled on in conversation 1. Anything worth keeping needs to go into knowledge or instructions instead of sitting in a conversation. Turning on Claude's memory in settings does give each project its own separate memory space, but memory is a different mechanism from project knowledge. Play it safe and put anything you truly need to keep into knowledge or instructions. This part varies by service: ChatGPT works the opposite way, and conversations within the same project do reference each other.

There's a genuine perk here too. Claude doesn't burn through your usage limit again each time it references material already in a project, because uploaded material gets cached. That's according to Claude's usage guidance documentation. The bigger the document and the more often you reference it, the more you gain from putting it in a project instead of attaching it fresh in every conversation.

[Hands-on] Building a client-report project

In the previous lesson, you got a monthly content report skeleton back as an artifact. At the time, you had to spell out the section order and the table columns in the prompt every single time. This time, you'll move that explanation into the room itself, permanently, then request the same skeleton again and compare what changes. A free Claude account works fine for this. Projects are available on every plan, and free accounts can create up to 5.

1. Start by creating the room. Click into Projects in the left sidebar, or go straight to claude.ai/projects for the same screen. Click New project in the top right and give it a name. The name is just a signpost for finding the room later, not something the AI reads, so anything easy for you to recognize works, something like "Client A Monthly Report."

2. Upload the material you've been explaining every time. Click the plus button under project knowledge on the right and add your files. For this exercise, upload two kinds: one or two of last month's reports, and a guide document with your brand naming rules. The past reports serve as your format template. The brand guide prevents naming mistakes, so you don't have to correct the same thing (whether to write the client's name in English or Korean, say) every single time.

3. Write the instructions. Click into project instructions, write down how Claude should behave in this room, and save. What to actually write is coming up in the next section with a real example, so it's fine to leave this blank for now and come back to it.

4. Send the same request again. Open a new conversation inside the room and send a shortened version of the request from the previous lesson.

Make me a skeleton for this month's content performance report.

In the previous lesson, you had to spell out the section order and table columns in the format field every time. Inside the room, you can drop that explanation entirely and still get a skeleton that follows the structure of your past reports. The output still lands in an artifact window, same as before. The room supplies the material, and the artifact shows the result.

How to write good project instructions

First attempts at writing instructions usually fail in one of two directions: a single line like "do a good job," or a two-page rulebook at the other extreme. So what should actually go in there? Google's official documentation recommends four elements for writing Gem instructions, and they make the cleanest possible skeleton: Persona, Task, Context, and Format. The names come from Gemini's documentation, but they carry over directly to Claude's project instructions and ChatGPT's project instructions.

• Persona: What role to answer in and how.
• Task: What this room is supposed to produce.
• Context: Give as much background as you can, including which uploaded file counts as the standard to follow.
• Format: Spell out the structure you want, specifically.

For a client-report room, it looks like this. Copy it as-is and swap in your own names.

Role: You write B2B content marketing reports. Keep sentences short and the tone professional.
Task: When asked, write the client's monthly content performance report.
Source material: The past reports in project knowledge set the format. Follow the brand guide for the client name and product names.
Format: A 5-line summary, a metrics table, then 3 suggestions for next month, in that order. Numbers go in the table only, don't repeat them in the body text.
Check: Don't make up numbers that aren't in the source material. Ask me instead.

What about a list of don'ts? Keep it to essentials only, like that last line above. OpenAI's guide to writing GPT instructions advises spelling out what you want specifically rather than piling up a long list of don'ts. That advice was written for GPTs, but it applies directly to project instructions too. "Don't sound stiff" does far less work than "keep sentences under 40 characters, professional tone." Cap your list of don'ts at two lines, and spend the extra line on the format section instead.

Account-wide instructions vs. project instructions: what goes where

This can look like it overlaps with the custom instructions you set up in the previous track. There's one rule: anything about you as a person goes in the account, anything about the job goes in the project.

Claude manages account-wide instructions under Instructions for Claude in settings, which you reach by clicking your account initials in the bottom left. Claude's personalization documentation draws the scope line this way: account-wide instructions apply to every conversation, project instructions apply only to conversations inside that project. The fact that you're a marketer belongs in the account. The table structure for this client's report belongs in the project.

ChatGPT goes a step further and sets a priority order. According to OpenAI's official documentation, project instructions apply only within that project and override your account-wide custom instructions. When the two conflict, the project wins. Say your account instructions read "always summarize in bullet points." If a report room needs prose instead, just write that into the project instructions and it takes precedence.

How are artifacts and projects different

This question comes up constantly in search. In one sentence: an artifact is the window where output shows up, and a project is the room where material and instructions live. An artifact is a document, table, or mini-app Claude built, opened in a separate window next to the conversation. A project is where you keep the material that goes into producing that output. One is what comes out, the other is what goes in, so they're not actually competing with each other. What an artifact is got covered in the previous lesson, so this section is just about how the two relate.

Have a conversation inside a room, and the output comes out as an artifact window. Think of using the two together as the default. Where they're stored is different, though. An artifact created inside a conversation doesn't automatically show up in the sidebar's artifact list. Claude's official documentation states that you need to open the artifact and click Publish to add it to that list. The room is where your raw material lives; the artifact list is a separate shelf for output you deliberately kept.

ChatGPT Projects and Gemini Gems

All three services run the same structure under slightly different names. Swap the labels, and everything you just learned carries straight over.

ChatGPT projects are available on every plan, Free included. You write instructions from Project settings, under the three-dot menu in the top right of a project.

ChatGPT has one device Claude doesn't: project-only memory. Choose it, and conversations in that room stop referencing memory saved before it existed. They also stop referencing anything from conversations outside the room. Conversations within the same room still reference each other as normal. It's a useful partition when you want to keep sensitive client work separate from your personal conversations. You can choose this when creating a project, and you can change it later on a project you already built. According to OpenAI's official documentation, you pick between general memory and project-only memory from the memory option in the same project settings screen where you wrote your instructions, then save. Changes can take a few hours to take effect. One catch: a shared project automatically switches to project-only memory and can't be switched back to general memory. Using project memory at all requires both "reference saved memories" and "reference chat history" turned on in settings.

Gemini's Gems are free to use. Google's documentation on plan-by-plan features lists Gems as available at every tier in a comparison table, from no Google AI plan at all through Plus, Pro, and Ultra. Creating, editing, and deleting all happen at gemini.google.com in a browser. This works from a phone or tablet too, but Google's documented mobile process is still the browser at gemini.google.com, not the Gemini app. Once a Gem is created, it shows up in the mobile app and in the Gemini side panel inside Google Workspace. In the Gem builder, this field is simply labeled Instructions, with file uploads handled by the Knowledge section right below it.

This is also where the official documentation is friendliest about how to actually write instructions. The four elements used as the skeleton earlier in this lesson come straight from that document. If you're stuck on where to start, it includes three full example instructions: a brainstorming tool, a coding partner, and a writing editor, each showing exactly how the four elements got filled in. One more note: using Gems with a personal Google account requires being 13 or older.

When a project won't show up or won't open

This one comes up constantly in search too. Most of the time, nothing was deleted, it's either sitting somewhere else or the page just failed to load. One thing worth knowing up front: OpenAI's published checklist is written for when chat history goes missing, not specifically for projects. There's no separate official procedure just for recovering a project. What follows adapts that same checklist to a project situation. Go through it in order, top to bottom.

1. Check your account first. OpenAI's help center says the first thing to check when chat history looks empty is whether you're signed into the right account and workspace. If you juggle a personal account and a work account, this is where the mystery usually ends.
2. Check the archive. In ChatGPT, check Archived chats under Settings > Data controls. It might be archived rather than deleted. Claude also has an archive feature for projects: an archived project drops to the bottom of the list and shows up under the Archived tab. Its knowledge and permissions stay exactly as they were, and unarchiving restores everything. Claude's documentation states that an archived project can't even be deleted, you have to unarchive it first.
3. Check your chat history setting. If ChatGPT's chat history setting is turned off, the list can look empty.
4. Refresh and sign back in. Reloading the page, or signing out and back in, forces the list to reload.
5. Check service status. An ongoing outage listed at status.openai.com can make history disappear. OpenAI's troubleshooting documentation also recommends turning off browser extensions, VPNs, and secure DNS tools, trying an incognito window or a different browser, and clearing your browser cache and cookies.
6. Still missing? Export your data. Requesting a data export from Settings > Data controls tells you exactly what's actually still on your account.

If you use a work account, there's one more thing to check. If you just merged a personal workspace into a company workspace, conversations and GPTs can take about 24 hours to show up while reindexing catches up. Don't panic if things are missing on the day of the move itself.

I learned this pattern first through ChatGPT's GPTs. I built three, and looking back, all three came down to the same split: knowledge on one side, instructions on the other.

Marketing copy came first. I put the output format in the instructions, and when I actually used it, all I had to paste in was one URL plus a line like "Instagram caption, aimed at women in their twenties." That was enough to get a usable draft every time, no need to re-explain what our company does or what tone we use.

Studying English was second. The instructions were simple: translate anything I wrote in Korean into English, and if I wrote in English, treat it as my own draft and correct it. I specified the standard for corrections too: conversational, current American English, phrasing people in their twenties and thirties actually use. And wherever the grammar or word choice was off, I had it flag the spot and explain in Korean, since the whole point was learning.

BigQuery was third, and here the knowledge side carried more weight. I uploaded the full table structure up front, so all I had to write was what I wanted to see, and it wrote the query. The part where I used to paste table names and column lists by hand every time just disappeared.

None of the three worked on the first try. My first draft of the instructions never got the output I wanted, and most of the time the problem was my own explanation being too thin. So I'd open a separate chat, rewrite the instructions there, and paste the revised version back in, over and over. That's when I learned instructions aren't something you write once and leave alone. You write them, see what comes out wrong, and rewrite. That holds whether it's a Claude project or a Gem.