Claude Artifacts and Projects: 6 Common Mistakes and a Checklist to Catch Them

"I got as far as building a project. Am I actually using it right, though?"

Six mistakes account for most of the trouble in this stretch: a setting that keeps the artifact window from opening, an artifact that never shows up in your list because you forgot to publish it, presentation slides you expect an artifact to produce, a project name and description Claude never actually reads, a project stuffed with files but no instructions, and a project you rebuild from scratch because you assumed it vanished. For most of these six, the cause is spelled out in the official docs, and the fix is one line. If you loaded up files and instructions in the previous lesson, this one is where you confirm the room and the window actually work the way you think they do. If you came here because an artifact or a project isn't behaving, this lesson alone should cover it. As of August 2026.

7 things to check before you close out this track

Let's start with the checklist itself. You've done each of these once already, somewhere in the last three lessons. Reading about something and actually having done it are different things, so just mark each one yes or no, honestly.

1. I scrolled the sidebar all the way down and noted how many of the four steps, Grab, Organize, Connect, and Delegate, actually show up on my screen.
2. I asked for a report skeleton and watched an artifact window open on the right side of the chat.
3. I sent a follow-up edit to create version 2, then used the version selector to switch between version 1 and version 2.
4. I took a finished artifact out of the conversation, by publishing it, copying it, or downloading it.
5. I created a project and added at least one file to its project knowledge.
6. I wrote four lines into the project instructions: role, task, source material, and format.
7. I sent a plain one-line request in that project, and the answer reflected the format or terminology from the files I'd uploaded.

The 6 most common mistakes in this stretch

Where the misses land is pretty predictable. Here are all six, symptom, cause, and fix. The fine print that got a passing mention earlier in the track is exactly what trips people up here.

Mistake 1: A setting that keeps the artifact window from opening

You clearly asked for a document, and the answer just spills straight into the chat instead. No window opens on the right, so the output gets folded right back into the conversation.

Check one setting first. It's on by default across nearly every plan, including Free, so this usually isn't the problem, but if someone switched it off, that's the whole story. Claude's help center states plainly that Claude no longer supports artifacts once Code execution and file creation is turned off under Settings > Capabilities. Go check that toggle first. On an organization account, only the owner can flip it in the org's settings. If the toggle is already on and the window still won't open, your request probably didn't clear the bar. The documented threshold is content over 15 lines that's substantial and self-contained enough to stand on its own outside the conversation. Ask for a one-paragraph blurb as an artifact, and the window may just not show up.

Mistake 2: Why a finished artifact isn't in your list

You go looking for yesterday's report skeleton under Artifacts in the sidebar, and the list is empty. That's exactly the situation behind searches like "claude artifacts missing."

It wasn't deleted. Claude's documentation notes that artifacts created inside a conversation don't automatically appear in the sidebar. To land there, you open the artifact and click Publish. Worth checking your plan here too: the Publish button only exists on Free, Pro, and Max. On Team and Enterprise, the same spot is labeled Share instead, and sharing on an org account stays inside the organization rather than going public. So if it's something you'll want to reuse later, publish it while you're already there polishing it. One catch, though: Publish isn't a button you tap and untap casually. Publishing exports whatever version is currently open, and once you unpublish, you can't republish that same artifact, you'd have to create a new one to bring it back. Before you click, make sure the version selector is actually sitting on the version you mean to export.

Mistake 3: Expecting an artifact to hand you a slide deck

Common enough that "claude artifact ppt" shows up in autocomplete on its own. You want a report turned into slides to attach for a client, and no file ever appears, so you go hunting for it.

Artifacts won't do this. The documentation's six content types are documents, code, single-page websites, SVG images, diagrams and flowcharts, and interactive React components often called mini-apps. Slides aren't on that list. File creation is the feature that actually does this: Claude's documentation states outright that it can produce Excel (.xlsx), PowerPoint (.pptx), Word (.docx), and PDF files, up to 30MB each. The two features get confused because they share one switch, both turn on through the same Settings > Capabilities toggle from Mistake 1. Use this rule of thumb: something you'll keep tweaking next to the conversation is an artifact; something you'll attach to an email is file creation.

Mistake 4: Putting information where the AI never reads it

You type "monthly report for Client A, keep sentences short" into a project's description field and wait for the answers to change. Or you assume the table format you agreed on yesterday in Claude will just carry over into a new conversation you open today.

Both assumptions run backward from what the documentation actually says. Claude's docs state that a project's name and description are just labels Claude doesn't access, and that conversations inside the same project don't share context with each other unless something is added to project knowledge. The name and description are a sign that helps you find the right room; each conversation is a one-off event that happened inside it. ChatGPT works the opposite way here, conversations in the same project do reference each other. Claude does generate a separate memory per project on paid plans that use memory, and some of it sticks, but that's a different mechanism from project knowledge, and you don't get to choose what survives. Anything you actually need to keep belongs in instructions or knowledge, not left sitting in a conversation. If you nailed down a table format on your third conversation, updating that one line in your instructions saves next month's you a headache.

Mistake 5: A project stacked with files and no instructions

You've uploaded ten files, and the answers you get inside the project don't read any different from the answers you'd get outside it. Piling on more files doesn't fix it.

Knowledge and instructions play different roles. Knowledge is the ingredients; instructions are the recipe. Claude's documentation notes that project instructions let you set how Claude responds within a given project. Stack up ingredients with no recipe, and the standard shifts every time. Upload two past reports and a brand guide without noting which one sets the format, and Claude just picks whichever one it feels like. Four lines are plenty: what role to answer in, what this project is for, which uploaded file is the standard to follow, and what structure the output should take. This maps onto the four elements Google recommends for writing Gem instructions, Persona, Task, Context, and Format, carried over to a project, and the same four lines work just as well in Claude's or ChatGPT's project instructions.

Mistake 6: Recreating a project you assume vanished

"chatgpt can't load project" and "chatgpt project disappeared" both show up in autocomplete side by side. The room's nowhere to be found, so in a hurry, people rebuild it under the same name.

Check before you rebuild. Most of the time it wasn't deleted, it's sitting somewhere else, or the page just failed to load it. Start by confirming you're signed into the right account and workspace, then check whether it's archived. Where you look differs by service. ChatGPT's documentation doesn't spell out a dedicated path for archiving a project itself, so you end up checking Archived Chats under Settings > Data controls; Claude keeps archived projects in the archived projects tab on the Projects page. Claude's documentation states that archiving preserves members, permission levels, and project knowledge exactly as they were. An archived project can't even be deleted; you have to unarchive it first. None of that resembles something that's actually gone. Lesson 3 in this track walks through ChatGPT's check in six steps, follow that same order. Rebuilding from scratch means refilling instructions and knowledge from zero. The five minutes it takes to check is much cheaper.

When do I get to learn Code and Cowork

Where these two live and what they actually are got covered in Lesson 1. What's worth confirming today is just the order. Code and Cowork both belong to Step 4, Delegate. Delegating isn't a step where you check the result, it's a step where you hand over the whole process. Open either one before you have a feel for what's safe to hand off, and you'll have no basis for trusting what comes back. That's why Code is saved for the CLI track and Cowork for the advanced track after that. If there's still an item in the sidebar you haven't clicked, or a mode in the input box you haven't tried, that doesn't mean you're missing something.

What you upload to a project is a snapshot, not a live copy

One more thing before this track wraps up. Think back to that client-report project from earlier. You uploaded last month's report and the brand guide to project knowledge. Those files are frozen exactly the way they looked the moment you uploaded them.

What happens next month? The performance numbers get updated, and the brand guide picks up a line for a new product. The project itself doesn't change on its own. You'd have to download the new files and re-upload them yourself. Manageable for one client. Multiply that by five clients with three files each, and you're re-uploading fifteen times a month. You built the project to reuse material, and hauling files around by hand is exactly the work that's still sitting there.

The documentation already has an answer for this. Claude states that Google Docs pulled in from Google Drive stay synced with the latest version in Drive. Instead of carrying copies back and forth, you keep the original linked. Two conditions come attached, though. That sync promise only covers Google Docs; spreadsheets and PDFs are read once and that's it. And this path only works in private projects, it's disabled in projects shared with an organization. What makes that link possible is a connector. A connector doesn't read a file you uploaded, it reaches into your own service and pulls whatever it needs. What it can see is exactly what you already have permission to see there. If you can't open a file, Claude can't either.

That's why Lesson 1 placed Connect at Step 3 and pushed it outside this track. Not because it's technically harder to operate, it's that it takes judgment about what to allow and when to pull it back. If you checked off five or more on that checklist, you're ready to make that call.

I held out inside the web tools for a long time too. But the moment I hit a wall wasn't about hauling files around. It was about waiting.

I'd type out a long explanation in chat, wait for the Artifact to catch up, and get back something completely different from what I asked for. So I'd explain again and wait again. After a few rounds of that, it started to feel like my actual job was waiting, not building. The longer my explanation got, the longer the wait got too, and a longer explanation didn't even guarantee a better result.

That's the edge of the web workspace. Creating a room and loading it with files works fine. The problem is the room isn't actually connected to your real files, so you have to close the gap with words, every single time. The next track cuts that explaining down.