Live Artifacts
Dashboards that go and check for themselves — built by Claude, inside Cowork, on your desktop.
Kevin Lenida
Welcome. Today is about one narrow feature that changes a habit most of us have without noticing: asking Claude the same question over and over. By the end you will have built one of these yourself. Everything here is accurate as of August 2026 — this feature is young and the app updates often, so if a menu has moved by the time you watch this, the shape of it still holds.
You have asked Claude the same question every Monday for a month.
Let this sit. Ask for hands: who has re-asked Claude something this week that they also asked last week? Open tasks, what is in my inbox, what did the team ship, what are competitors doing. Every one of those is a question whose answer expires. Today we stop re-asking.
Ground zero
Thirty seconds on Cowork
Cowork is a mode, not an app — pick it in the message box, next to Chat.Chat talks. Cowork works. It runs multi-step jobs on your files and connected apps.Desktop is the full version — it is the only one that touches your own computer.
Live artifacts live entirely inside Cowork. Nothing today applies to Chat.
Quick level-set so nobody is lost. Cowork is the second mode inside the Claude app — same message box, you just pick Cowork instead of Chat before you type. Chat is for thinking; Cowork is for handing over a job that takes many steps and touches real things: your files, your Gmail, your project tracker. Cowork also runs on web and mobile now, but desktop is the complete version because only desktop can reach the files on your machine. That distinction is going to matter a lot later.
support.claude.com/en/articles/13345190-get-started-with-claude-cowork
01
The Idea
What a live artifact is, and what "live" honestly means.
The Idea
Two things share the word "artifact"
Chat artifact
Made inside one conversation
The data stops when Claude builds it
To update it, you go find the chat
Live artifact
Has its own place in the sidebar
Re-reads your apps on open
You just open it
This is the single most important slide for avoiding confusion, because both things are called artifacts. A chat artifact is a photograph — accurate the second it was taken, and quietly wrong forever after. A live artifact is a window. Same word, completely different behaviour. In the app you will see live ones labelled "Cowork" so you can tell them apart at a glance.
The Idea
How live is "live", really?
FROZEN
Chat artifact
Built once. Never moves.
FRESH ON OPEN
Live artifact
It re-asks when opened.
REALTIME
A trading screen
Changes as you watch.
Ask for the middle one and you can still get the left one.
Live does not mean realtime — nothing streams, nothing pushes. It means that when you open the page, it goes and asks your connected apps what is true now, and redraws. You will actually see it happen: grey placeholder bars for a second while it fetches, then the content. Now the part that matters more than the definition, because I got caught by it for two hours while making this deck. You can very easily end up with a page that only LOOKS live — Claude builds a static snapshot, saves it to your Artifacts sidebar, and the header still says "Data updated just now" every time you open it. I reloaded a fake one four different ways and it never changed a word. The difference is where you build it, and that is the next section. One more practical thing, learned the hard way: a real one does not necessarily fetch just because you opened it. There is a cache in front, Anthropic does not publish how long it lasts, and in my testing opening the page five minutes after the last fetch served me the old data while the header counted up "5 minutes ago". If you need to be certain you are looking at now, press Reload. That is the only deterministic path, and it is what I will be pressing in the demos.
The Idea
Four things you actually get
Its own front door — not buried inside a chat.Current data — on open it re-reads your connected apps.A history — every change is a version you can restore.Free to open — a refresh runs code, not Claude.
A dashboard, a tracker, a morning brief — any page you would otherwise rebuild.
Four properties, each solving a real annoyance. First: you never have to remember which conversation made the thing — it has its own home in the sidebar. Second: it is current, because opening it sends it off to ask your apps again. Third, the underrated one: version history. Ask for a change you regret and you can restore the previous version instead of describing your way back to it. The fourth is the one I would put money on them not expecting, so slow down here. Opening the page costs you nothing. The page holds ordinary code — it calls your connectors directly and renders the result, and no model reads any of it. So building the thing costs usage once, and every open after that is free, forever. Put that against the habit we started with: asking Claude the same question every Monday costs you something every Monday. Build it once and stop paying for the same answer. Someone will push back here, because MCP stands for Model Context Protocol and it sounds like a model must be involved. Have this ready: the name says who it was designed for, not who has to be present. It is a wire protocol — the caller asks for a tool by name with some arguments, the server runs it and returns JSON. No model sits inside that exchange. Normally a model is the one calling, because choosing which tool and which arguments is the hard part — but the call itself is just a request and a response, like any API you could run by hand. So: Claude decides which calls to make once, and writes them down. Reloading just runs what it wrote.
support.claude.com/en/articles/14729249-use-live-artifacts-in-claude-cowork
Notice the second time you ask.
This is the trigger I want them carrying for the rest of the hour, which is why it sits here and not at the end. You do not have to judge whether something is worth building, or learn a rule. You just have to notice a feeling — I have answered this before. The second time is the signal. A question you ask once is just a question; answer it in Chat and move on. Give them a second way in as well, because plenty of people never notice their own repetition but do know when they want a dashboard. So: the moment you catch yourself thinking "I wish I had a dashboard for this" — that thought is a live artifact. Dashboard is the word they already have for it, and it is worth handing them, because otherwise they will hear "artifact" and think it means something exotic. What kind of thing to build is the next slide.
The Idea
Asked twice — but which one?
Important honesty slide, and it protects the line you just heard. "Asked it twice" is the right trigger, but it does not always point here — Claude has three different answers to repetition, and picking the wrong one is why people end up disappointed. Ask what exactly is repeating. If a QUESTION repeats — what is on my plate, what did the team ship, what are competitors doing — you want a view you can open, and that is a live artifact. If a JOB repeats — summarise last week every Friday, tidy that folder every night — you do not want to open anything, you want the finished thing waiting for you. That is a scheduled task: you describe it once, it runs on its cadence in the cloud, and it keeps running with your laptop shut. If a METHOD repeats — the way you format a report, the checks you always run, your house style — that is a skill: procedural knowledge Claude loads by itself when the work calls for it. Two things worth saying out loud. They combine — a scheduled task can use your skills and your connectors, so this is not a competition. And scheduled tasks are available on every paid plan and run in the cloud, which makes them the answer when the thing you want is not a page but an outcome.
support.claude.com/en/articles/13854387-schedule-recurring-tasks-in-claude-cowork
02
Demo — Morning Brief
Built from nothing, on screen, using Google Calendar and Gmail.
Copy the demo prompt
Build me a live artifact: a morning brief. Today's events from Google Calendar, and unread important mail from Gmail. Put whatever needs me first at the top.
Prompt: "Build me a live artifact: a morning brief. Today's events from Google Calendar, and unread important mail from Gmail. Put whatever needs me first at the top." Switch to the app. Goal is about five minutes. Narrate the ask out loud as you type it — especially the part where you name Google Calendar and Gmail by name, because that is the habit the next section teaches. Let them watch it work; do not fill every silence. When it opens, read one row out loud and point out that this row did not exist when you started talking.
03
How To Build One
Where to start it, and the one sentence that decides whether it works.
How To Build
Build it on your own machine
Start Cowork locally — the session must run on your computer, not in the cloud.Say the word "live" — ask for a live artifact, not just a page.Name your sources — the connectors it should read, by name.
Then check it: press Reload. A live one flickers grey while it fetches. A snapshot never moves.
This is the slide I could not have written yesterday, and it is the difference between a page that works and one that quietly lies to you. Where the session runs decides everything. A Cowork session on your own machine can hand the artifact access to your connectors. A session in the cloud cannot — artifacts made there are born with no connector access, so whatever Claude builds can only be a snapshot of what it read while you were talking. Here is the sharp edge. The New artifact button does not choose for you: it inherits whatever your Cowork session was last set to, cloud or local. Nothing on screen tells you that this matters, nothing warns you, and it does not switch you to local even though a live artifact cannot work any other way. I built five dead pages before I worked that out. So check the session location before you ask, every time — and if you take one operational habit from today, make it that one.
How To Build
Name your sources out loud
You
Build me a live artifact: a morning brief. Today's events from
Google Calendar , and unread important mail from
Gmail . Put whatever needs me first at the top.
Claude
Done — it is in your Artifacts sidebar as a live artifact. It re-checks both every time you open it.
Copy this prompt
Build me a live artifact: a morning brief. Today's events from Google Calendar, and unread important mail from Gmail. Put whatever needs me first at the top.
This is the whole skill, and it is one rule: say where the data comes from. "Build me a dashboard of my work" produces a beautiful page full of nothing, because Claude had to guess which apps you meant. Naming Google Calendar and Gmail tells it exactly which doors to knock on, and those named connectors become the permissions the artifact carries from then on. Second thing worth copying from this prompt: the last sentence tells it how to sort. Say what "good" looks like, not just what data to fetch.
How To Build
Four ways to build a bad one
Vague sources — "my tasks" gets you a guess; "my Linear issues" gets you data.Data it cannot reach — if no connector or folder can see it, no page can show it.A connector that can change things — it will act without asking.Do not assume it is live — check that it reloads. Snapshots lie.
These are the three failures I'd bet money on seeing in the hands-on later. The first is fixable by rereading the previous slide. The second catches people out — you cannot build a live view of something living only in your head, or in an app you have not connected. Check the connector exists before you describe the dream. The third is the one that stops being an inconvenience and starts being a problem, and it gets its own section.
04
Demo — Project Tracker
Build it, change the source underneath it, watch it catch up.
Copy the demo prompt
Build me a live artifact: a project tracker from my Linear issues. Group them by project, show status and who it is assigned to, and put anything blocked or overdue at the top.
Prompt: "Build me a live artifact: a project tracker from my Linear issues. Group them by project, show status and who it is assigned to, and put anything blocked or overdue at the top." This is the demo that sells the feature, so do not rush it. Build the tracker, and set its refresh to on-demand when it asks. Then, with the artifact on screen, switch to Linear and add an issue in front of them. Come back, go to Scheduled, and run the task. Watch the page rewrite itself with the new row. Say the two things that matter: nobody had a conversation with Claude to make that happen, and the page did not go and look — something rebuilt it. If the room only remembers one screen from today, make it this one.
05
The Fine Print
Three facts that decide whether this is safe to keep.
A live artifact uses your connectors without asking .
Slow down. This is the one thing in the hour I'd be upset if they missed. A live artifact can only use connectors you approved when it was built or last updated — that part is fine. But within that set it does not prompt you. It will not stop and ask "may I open Gmail". And it skips the prompt even in a session mode that would normally need your approval for every tool call. Anthropic states this plainly as a current limitation, in their own documentation.
The Fine Print
The fair version of the risk
It holds no passwords — the page never sees a token.But it acts unprompted — no confirmation step, ever.Do not ask for actions — no buttons that send, close or delete.Judge the worst case — you approve a whole connector, not one safe corner of it.
Both halves, because either one alone is misleading. The reassuring half: your access tokens never sit inside the artifact — the calls are made server-side, so there is no credential in the page to leak. The uncomfortable half: no confirmation step at any point. Here is the mechanism, and it is worth saying slowly. A live artifact is a page of code Claude wrote for you. It runs when you open it. Whatever calls are in that code fire, against the connectors you approved when you built it. So the real question is not "will it write" — it is "did the code include a write call". That is mostly under your control, which is what the third point is about: if you never ask for a button that sends or closes or deletes, you do not get one. The fourth point is the part people miss. The grant is for each connector, not for each action — there is no fetch-only Gmail. Anthropic's own wording of this limitation warns you to take care with connectors that can change your data, so writes are absolutely possible, and sometimes they are exactly what you wanted. Just decide that on purpose.
The Fine Print
It lives on one computer
Desktop only — Mac, Windows, Linux beta. Paid plans.Not on web or mobile — they do not even appear in the list.It does not follow you — switch laptops, they stay behind.Claude reads local files once — no folder connector, so they do not refresh.
Cowork sessions and files follow you between devices. These do not.
Worth explaining the why, because it is not an arbitrary restriction. The artifact is stored on your machine rather than in your Claude account — that is also exactly what lets it read a folder on your disk. Local storage is the price of local access. Everything else in Cowork now runs in the cloud and follows you around; this one thing stays put. So do not build the team's critical dashboard on the laptop you are about to replace.
The Fine Print
How To Share It
Team and Enterprise only — on Pro and Max you cannot share or publish at all.Inside your organisation — a link anyone in the org can open. Never public.It runs on their access — sections they cannot reach show an error, not your data.
Check which plan your people are on before promising this. If you are on Pro or Max, live artifacts are yours alone — full stop. "Organisation" here means the Team or Enterprise workspace itself, the account your seats and admin settings live under; that is the boundary, and it is why individual plans have no sharing at all. There is a Share button in the artifact header that copies an organisation-wide link — no picking individual recipients, no public link, and they will need Claude Desktop to open it. Now the part that surprises people, so say it explicitly. You are sharing the recipe, not the data. The artifact is a set of instructions — fetch this, lay it out like that — and when a colleague opens it those instructions run against THEIR connectors. They see their meetings, their issues, not yours. Five people open the same artifact and five different pages render. Two consequences. First, nobody can see data they were not already entitled to; the worst case is an error message where a section would be, never a leak. Second, and this is the practical one: a shared team dashboard only works if everyone has the same connectors set up. If you are the only person with the analytics tool connected, your teammates get an error where your chart was. Check that before you promise anyone a dashboard.
support.claude.com/en/articles/14729249-use-live-artifacts-in-claude-cowork
06
Demo — Local Files
A folder on my disk, rolled up into a page — and then rolled back.
no connector — just files Copy the demo prompt
Build me a live artifact: a summary page from the documents in my Client Notes folder. One row for each document with its key points and its date, newest first.
Prompt: "Build me a live artifact: a summary page from the documents in my Client Notes folder. One row for each document with its key points and its date, newest first." — swap in the real folder name before the session. This is the one demo where you pick a working folder, so say that out loud as you do it. While it builds: no upload, no copy-paste, no cloud — it is reading a folder that has never left this laptop, including a PDF and a spreadsheet. Be straight about the limitation when the page appears, because it is the honest end of the story: live means connectors, and there is no connector for a folder, so this page is a snapshot. Add a note tomorrow and it will not appear until something rebuilds the page. That is the trade — local files are the one thing nothing else can read, and the price is that they do not stay live on their own. Then ask for one layout change, show the previous version was saved, and restore it.
Recap
What we covered
Asked it twice? A question you ask again is a live artifact.
Build it locally and name your sources, or you get a snapshot.
It acts on approved connectors without asking. Keep them read-only.
Three lines, matching the three things I said they would walk out able to do. Read them out, do not elaborate — elaboration here just dilutes it. If they only keep one, it is the first.
Anything you ask Claude twice should stop being a conversation .
Say it, then stop talking. This is the sentence I want surviving the walk to the car, and it is deliberately wider than today's topic. A repeating question becomes a live artifact — that is the hour they just spent. A repeating job becomes a scheduled task. A repeating method becomes a skill. All three replace the same bad habit: retyping something you have already asked. If they only ever build live artifacts, fine, they still won. But the instinct is the prize.
Your turn
Build one now
Think of a question you ask twice — or a dashboard you wish existed.
Check which of your connected apps holds that answer.
Start Cowork on your own machine — not in the cloud.
Ask for a live artifact, and name those apps in the sentence.
Press Reload and watch it fetch.
One rule while you build: read-only connectors only. Nothing that can send, pay, or delete.
Fifteen minutes, everyone builds. Circulate. The two problems you will be fixing over and over: prompts that do not name a source, and people trying to build over an app they never connected. Ask for two volunteers to share their screen at the end — seeing a peer's beats seeing mine.
Now go build the thing you keep re-asking for .
See you in the sidebar.
Last slide, and the only one that is not about the feature. Say something like: everything today came down to noticing one small habit — that you keep asking the same question. You do not need permission, a budget, or an engineer. You need a laptop you already own and a sentence you can already write. Then stop. Let the room clap or laugh at the confetti, and go straight into questions.