Claude Artifacts and File Creation: Excel, PowerPoint, Word, PDF and Interactive Visuals
The way to use an Excel report produced by Claude safely is to write the definition into the brief, ask for formulas and add a checks sheet that calculates the total two ways. Artifacts, available on every plan since 16 September 2026, temporary in-conversation visuals, creating xlsx, pptx, docx and pdf files in a sandbox, sharing options by plan, and Bookstore's monthly sales report.
Version note: This post covers artifacts, in-conversation visuals and file creation in Claude as of 15 November 2026; plan rules, sharing options and limits were checked the same day in the Claude Help Center (support.claude.com). This post is based on the official documentation. The sample CSV data is made-up Bookstore data for this post; it is not a real sales figure.
Guess First
Selin gave Claude the October sales CSV of Bookstore, the online bookstore used as the shared example across this site's series, and asked for an Excel report with a pivot table by category. The file looked fine; but Hakan from institutional sales said the net sales total at the bottom of the report did not match the figure he took from the accounting system. The same day, Selin had also shared an artifact with a summary of the report through "Anyone with the link".
Two questions: how does Selin find out whether the figure in Excel or the one in accounting is right, and how could she have made that easier from the start? Can someone who gets hold of the link open the artifact without signing in to Selin's account, and if so, what do they see? The answer to the second question is in the Help Center's artifact sharing article; the answer to the first is more a habit than a setting.
What an Artifact Is, and What Changed in 2026
According to the Help Center article "What are artifacts and how do I use them?" (https://support.claude.com/en/articles/17153992), an artifact is content Claude produces for you that you can edit, reuse and share with others; it opens beside the conversation. Documents, code snippets, single-page websites, images, diagrams, dashboards and small interactive tools fall into this scope. Claude creates an artifact when the content stands on its own, is usually longer than 15 lines and looks like something that will be edited later.
The most important change in 2026 is access. According to the release notes, since 16 September 2026 artifacts are available on every plan, including Free. The ready-made templates also depend on the plan: Docs, Slides and Design on all plans; Data on Pro and above; Sheets and Whiteboards in beta on Pro and above; Motion in beta on Team and Enterprise. All artifacts are collected in the Artifacts tab in the sidebar and can be used on the web, Desktop, Claude Code and the mobile apps; on mobile you can view and edit them, but sharing settings need the web or Desktop. Artifacts require code execution and file creation to be turned on in the Capabilities settings.
Artifacts created in conversations before 16 September 2026 keep working and can still be shared as "legacy" artifacts, but no new artifacts can be created in that form. Editing works in a few ways: asking Claude for a change, editing directly in template-based artifacts, or selecting text in a document or Markdown file and choosing "Edit with Claude". Docs can be exported as Word, PDF, Markdown and Google Docs, and Slides as PowerPoint, PDF and Google Slides.
In-Conversation Visuals Versus Artifacts
Since 12 March 2026 Claude can also draw interactive charts and diagrams inside the answer. According to the Help Center article "Custom visuals in chat and Cowork" (https://support.claude.com/en/articles/13979539), these visuals are built with HTML, can be clicked and can be changed through the conversation. The feature is in beta; it works on the web and Desktop and does not show on iOS and Android. Claude decides on its own when a visual would help, but you can also ask explicitly with "show this as a chart"; the article notes that Opus gives the best results on complex visuals.
The difference from an artifact is persistence. In-conversation visuals are temporary; they are not saved separately as the conversation moves on. If you want to keep one, you can copy it as an image, download it as SVG or HTML, or save it as an artifact. For visualizing something quickly in a meeting, an in-conversation visual; for something to share with the team and keep updating, an artifact.
File Creation: Excel, PowerPoint, Word and PDF
Besides artifacts, Claude produces real files too. According to the Help Center article "Create and edit files with Claude" (https://support.claude.com/en/articles/12111783), Claude can create Excel (.xlsx), PowerPoint (.pptx), Word (.docx) and PDF files; it can analyze CSV and similar data files you upload and produce charts as PNG, and it can write plain text files such as .csv, .md and .html. It does this by running Python or JavaScript code in an isolated sandbox container inside claude.ai. For Office files there are ready-made skills from Anthropic, and Claude loads them on its own when needed.
The feature depends on the "Create and edit files" setting under Settings > Capabilities ("Code execution and file creation" in the older layout); on Team and Enterprise the organization owner manages it from Organization settings > Capabilities. The per-file limit is 30 MB for both upload and download. Created files appear in the conversation and can be downloaded or saved to Google Drive; in the unified experience the "Storage folder" setting in Settings shows where files are saved, and on Desktop the "Trusted folders" list shows which local folders Claude can reach. As we saw in the first post, creating files uses more than normal chat.
The sandbox's internet access is a setting too. On individual plans network egress is on and packages can be installed from approved sources. On Team the default is package managers only (such as npm, PyPI and GitHub), on Enterprise the default is off; owners can add specific domains to an allowlist or open all domains. The article calls this last option the riskiest; we see why in the In Production section.
Sharing: Who Sees What
According to the Help Center article "Share artifacts" (https://support.claude.com/en/articles/9547008), artifacts always start private to you. What comes next depends on the plan:
| Plan | Access options | Note |
|---|---|---|
| Free | "Publish" on legacy artifacts | A published artifact is open to anyone with the link |
| Pro, Max, Ultra | "Only you" or "Anyone with the link"; inviting people by e-mail (beta) | "Only you" stops sharing |
| Team, Enterprise | "Only people invited", "Anyone at" the organization or "Anyone with the link" | Sharing with groups on Enterprise; outside sharing depends on the admin's "External sharing" setting |
An invited person gets one of three permissions: "Can view" can read the artifact and its comments, "Commenter" can also comment and download files, "Can edit" can make changes as well. The answer to the second question in Guess First is here: most artifacts shared with "Anyone with the link" open without signing in; the exception is Docs, which require sign-in and are view-only. Someone coming from outside the organization or without signing in sees the artifact, but the parts that use connected apps and Claude do not work for them, and they cannot see comments. One more detail: everything saved in the artifact's shared storage is visible to everyone who opens the artifact.
Artifacts That Use Claude and Store Data
Artifacts do not have to be static pages. You can build small apps that call Claude; users sign in with their own Claude accounts, no API key is needed and usage comes out of each user's own plan limit, not the creator's. On Pro and above, on the web and Desktop, artifacts can also use apps you have connected (for example Google Calendar, Slack or your own custom connectors); each user approves the tools on first use and connects their own apps. In a shared artifact, the connected-app part runs on the viewer's own connection.
On Pro and above artifacts can also store data: personal storage private to each user, or shared storage visible to everyone. The default limit is 100 MB or 25,000 records, whichever is reached first; for uploaded files 20 MB per file, 5,000 files and 1 GB in total. That is enough for a small internal tracking tool; but it is not a product database.
Bookstore: Monthly Sales Report, Five Slides and a Stock Alert Page
Let us set up Selin's job again from the start, in a way that can be verified. She has the October sales export; its first rows look like this:
order_id,order_date,isbn,title,category,channel,quantity,unit_price_try,discount_try,status
10412,2026-10-01,9780441172719,Dune,Science Fiction,Retail,1,420.00,0.00,Delivered
10413,2026-10-01,9780141439587,Emma,Classics,Retail,2,260.00,26.00,Delivered
10414,2026-10-02,9780547928227,The Hobbit,Fantasy,Institutional,40,310.00,1240.00,Delivered
10415,2026-10-02,9780141439549,Middlemarch,Classics,Retail,1,350.00,0.00,Returned
10416,2026-10-03,9780441172719,Dune,Science Fiction,Retail,1,420.00,0.00,Cancelled
The brief for the report is written so that it shows where the figures come from:
Create an Excel file from sales-2026-10.csv (attached).
Sheet "Data": the raw rows, unchanged.
Sheet "Summary": net revenue by category and by channel as a pivot table.
Net revenue = quantity * unit_price_try - discount_try.
Count only status = Delivered. Show Returned and Cancelled totals in
separate columns, never inside net revenue.
Sheet "Checks":
- row count of Data, and row count per status;
- total net revenue calculated twice: from the pivot and with SUMIFS on Data.
Use Excel formulas that I can inspect, not pasted values.
Before you build it, list any rows you cannot parse and ask me what to do.
The most important parts of this brief are the "Checks" sheet and the formula requirement. Because the definition of net sales is written in the brief, the discussion with Hakan turns from "who is right" into "is the definition the same": does accounting subtract returns, where does it put the institutional discount? Because the total is calculated two different ways, a filter mistake in the pivot shows up as a difference on the Checks sheet. And because the formulas stay in the file, you can click any cell and read how it was calculated. The answer expected for this brief: a three-sheet Excel file, and, if there are rows it cannot parse, a question listing them first.
The presentation for management and the team's stock alert page come from the same data, in the same conversation:
Now create a 5-slide PowerPoint for the management meeting from the Summary sheet:
1) net revenue this month, 2) revenue by category, 3) Retail vs Institutional,
4) returns and cancellations, 5) three questions for the team.
Every number on a slide must come from a cell in the Summary or Checks sheet;
put the sheet and cell reference in the speaker notes.
Then create an artifact for the warehouse team: a page that lists books whose
stock is below the reorder level, using inventory-2026-10-31.csv (attached).
Sort by how far below the level each book is. No customer or price data.
Asking for cell references in the speaker notes makes it possible to trace every number on a slide back to Excel; a number copied from the wrong cell is a mistake that is hard to notice by looking at the slide. The stock alert page has a deliberate limit: the data comes from the export of 31 October, so the page is a snapshot of that day. If live stock data is wanted, the right way is to connect the page to Bookstore's own API through a connector; we see that in the eighth post. Keeping price and customer data off the page entirely also reduces the sharing risk from the start.
In Production
A created file is a draft. An Excel file with formulas is easier to verify than one with pasted values; but you still do the verification. Calculate the total in an independent way, check a few rows by hand, write the definition (what net sales includes) into the report.
Files and web pages can carry instructions. The Help Center warns that malicious instructions hidden in external files or websites can lead Claude to run untrusted code or send data from the conversation (prompts, projects, MCP data, Google integrations) to third parties. Its advice is to monitor Claude and stop it if you see it reaching data unexpectedly. Turning off network egress blocks data leaving the sandbox, but MCP connections can talk to the outside regardless of this setting. The defensive side of the topic is in the LLM Application Security post on layered defense against prompt injection.
There are built-in safeguards too: on Free, Pro, Max and Ultra, public sharing of conversations that contain file artifacts is disabled, sandboxes are not shared between users and a prompt injection classifier can stop execution. On team accounts, keeping network access at the "package managers" or allowlist level is a sensible default.
Common Pitfalls
Sharing without checking formulas and figures. A tidy-looking Excel file does not show that the figure is right. Write the definition, have the total calculated two ways, trace it back to the source.
Publishing sensitive data with "Anyone with the link". Someone with the link can see most artifacts without signing in, and everything in shared storage is visible to everyone. Share an artifact with customer or price data at most within the organization.
Taking an artifact for product code. An artifact is good for an internal tool or a draft; it is not Bookstore's customer-facing stock page. Product code is written in its own repository, with tests and review.
Thinking an in-conversation visual is permanent. It is not saved separately as the conversation moves on; save the visual you want to keep as an artifact or download it.
Choosing "Unpublish" on a legacy artifact without thinking. According to the Help Center, once a legacy artifact is unpublished it cannot be republished, and the data in its storage is permanently deleted.
A Small Experiment
★☆☆ Prepare a small CSV from your own data (without customer data) and ask Claude for a chart, first as an in-conversation visual, then as an artifact. Try to find both again after the conversation has moved on; where is each one?
★★☆ Run the Excel brief from this post with your own data. Do the two totals on the "Checks" sheet match? If not, find which rows the difference comes from and write down which definition is missing from the brief.
★★★ Write an artifact sharing rule for your team: which data can be shared at which access level, when "Anyone with the link" is forbidden, what never goes into shared storage, how legacy artifacts are handled. Review an existing artifact against this rule and fix its access.
For the Excel report, we provided the data. In the seventh post we move on to cases where Claude looks for the data itself: sourced research with web search and Research, checking citations, and verifying the breaking changes that affect Bookstore with official sources.