https://t.co/6Maq70xCSM

HeyGen on X: "https://t.co/6Maq70xCSM" / X

HeyGen

@HeyGen

$

Claude Cowork × HeyGen

3

1

21

13

A guide to using Claude Cowork with your connectors and the HeyGen MCP: what Cowork is now, how to wire Slack and HeyGen into it, and how to turn a week of your team's messages into an on-camera update without writing the script yourself.

💡 The core idea: Cowork turns what your team already wrote into something your audience will actually watch . Your Slack, docs and files are the raw material. Claude reads them, pitches you three videos, writes the one you pick, and hands it to HeyGen. You make two decisions: which story, and go.

00 · The whole thing, up front

If you read nothing else, this is the workflow from the video, start to finish. Connect HeyGen and Slack, switch them both on in the chat, and send this:

text

Read this week's Slack and make me a video using the HeyGen MCP.

Here is what comes back, in order:

• It reads. Claude opens the channel, counts the messages, groups them by topic, and tells you how many are worth a video.

• It pitches. Three video options, each with a one-line preview.

• You pick. Reply with the one you want: Let's go with option 1.

• It writes. A full script, built only from what was actually said in the channel.

• You say go. Okay, generate. That is the only step that spends HeyGen credits.

• HeyGen renders. Your Avatar, your voice, in your own HeyGen account.

That is three short replies for a finished update. The rest of this guide is how to make those three replies reliable, repeatable, and fast enough to run every single week.

01 · What Cowork is now

On September 16, 2026, Anthropic folded Cowork into Claude. There is no separate Cowork tab anymore: it is one Claude window, and Claude decides on its own whether a request needs a quick answer or real work across your files, folders and apps. When people say "Cowork" now, they mean that mode of Claude, the one that does the task instead of describing how you could do it.

• It works on real things. Folders you grant access to, the apps you connect, and a built-in browser it can drive itself. It opens the actual files, it doesn't ask you to paste them.

• It keeps working when you leave. Sessions run in the cloud. Close the laptop and pick up the same task from your phone.

• It is not a HeyGen feature. HeyGen plugs in through the same open MCP standard as Slack, Notion and Google Drive. Same HeyGen account, same Avatars, same credits.

02 · The loop

Four stages, and you only make two calls.

Stages one and three belong to Claude. Stages two and four are yours on purpose: you choose which story is worth telling, and you approve the render. Nothing spends a HeyGen credit until you say generate.

03 · What you need

• Claude on a paid plan (Pro, Max, Team or Enterprise). The desktop app gives you the most: folders, the built-in browser, and connectors in one place.

• A HeyGen account with an Avatar and a voice you are happy with. In the video, that's Keith with Keith's voice clone . Everything downstream inherits the quality of that Avatar, so build it first.

• A place your team already writes. Slack is the example here, but the same workflow works with Notion, Google Drive, Gmail, or a folder of docs.

• About five minutes , once.

04 · Connect HeyGen and Slack

• Add both connectors once. In Claude, open Customize → Connectors and add HeyGen and Slack.

• Sign in on HeyGen's own page. You grant scoped access and Claude holds a token you can revoke anytime.

• Switch them on in the chat. Click + in the message box and toggle HeyGen MCP and Slack on. The toggles apply per conversation, so only the tools you switch on are in play.

• Test the wire. Before you build anything, ask:

show me my HeyGen avatars and voices

If your Avatar names come back, the connection is live.

💡 No API key, ever . You are not pasting a secret into a config file or into the chat. You sign into HeyGen, grant access, and that's it.

💡 Golden nugget: switch on only what the job needs . Every live connector's tool list sits in Claude's context whether it gets used or not, and the more tools are switched on, the more often an agent reaches for the wrong one. For this job: HeyGen and Slack. Nothing else.

05 · The folder is the prompt

This is the part most people skip, and it's the reason the prompt in the video can be thirteen words long. The rules don't live in the prompt. They live in a folder.

Make a project folder, point Cowork at it with Work in a folder , and put one short instructions file inside. Claude reads it before every run, so every future prompt inherits it.

Here is a starting point you can copy and edit:

# Weekly update videos
Voice: plain, a little quick, like a voice note to the team. No hype.
Length: 25 seconds, about 70 words. Lead with the point.
Avatar: Keith. Voice: Keith's voice clone. 16:9.
Facts: only use what the source actually says. Never invent a number.
Always: pitch 3 options first.
Never: generate until I say "generate".
Tone: match the scripts in /past.

Why this works:

• You stop repeating yourself. Voice, length and Avatar are decided once, not in every message.

• It sounds more like you over time. Drop three to five scripts you liked into a /past folder. Claude matches their rhythm instead of guessing at "your style".

• The guardrails live in one place. "Never invent a number" and "never generate until I say so" are the two lines that keep this safe to run on a schedule.

💡 Cowork doesn't remember one session in the next on its own. Working inside a project and a folder is what carries your rules forward.

06 · The three prompts

1. Point it at the source and name the tool.

Naming both matters. "This week's Slack" tells it where to read, and "using the HeyGen MCP" removes any doubt about which connector should do the render. You'll see it work through the channel: reading, counting the messages, grouping by topic, then landing on three videos worth making.

1. Pick the story.

Let's go with option 1.

You are the editor now. It writes the full script under your folder rules. Read it before you move on. If something's off, just say so in plain words: "cut the second item", "make it under 20 seconds", "say ten percent, not 10%".

1. Approve the render.

Okay, generate.

This is the only message that spends credits, which is exactly why it should always be a separate message from the pick.

07 · Reading the HeyGen card

While it generates, a HeyGen MCP card shows each step as it happens. Learning to read it is the difference between trusting the process and refreshing nervously.

• Avatar and voice. It found the ones you named. If a name comes back you didn't expect, stop here and correct it. This is the cheapest moment to catch a mistake.

• Queued. The render job is in your HeyGen account, drawing on your normal credits. There is no separate MCP pricing.

• Processing, then completed. When it says completed, the video is sitting in your HeyGen library, ready to download, share, or translate.

08 · Write to the clock

Golden nugget: count words, not seconds. Asking for "a 25-second video" works sometimes. Asking for about 70 words works every time, because that's what actually decides the length.

Avatars read at roughly 2.5 to 3 words a second. In the video, Keith reads 67 words in 22.5 seconds. So:

• 10 seconds is 25 to 30 words: a hook and one point.

• 25 seconds is 65 to 75 words: the weekly update.

• 40 seconds is 100 to 120 words: a short explainer.

• 60 seconds is 150 to 180 words: room for a story.

Put the word count in your instructions file, next to the length. Also write numbers the way they should be said: "ten percent" reads better on camera than "10%".

09 · Scale it

Once the three prompts work, it's time to stop typing them.

Schedule it. Type /schedule and give it a standing job. Scheduled tasks run in the cloud, even with your laptop closed:

Every Friday at 2pm, read this week's #content and pitch me three video
options using my weekly-updates folder rules. Don't generate anything.
I'll pick one when I'm back.

You come back to three finished pitches instead of a blank page. Pick one, say generate, done.

Swap the source. The folder rules stay the same while the input changes: #content for the team update, a customer-wins channel for a client-facing clip, the product changelog for a release video. One setup, several shows.

Localize instead of re-recording. Once a video is generated, keep going in the same chat:

Translate this video into Spanish and French with HeyGen.

Same face, same performance, more languages, no new recording.

Save it as a Skill. When the workflow is stable, turn your instructions into a Skill so anyone on your team can run the exact same process with the exact same rules.

10 · How to use it effectively

• Keep the generate step manual. Automate the pitch, never the render. A scheduled job reads whatever is in the channel, including half-finished threads and messages that were never meant for a video. Your "generate" is the checkpoint.

• Check the facts before they go on camera. Claude only knows what's in the channel, and a thread that's still being debated can read like a decision. One wrong number in a published video costs more than a week of good ones.

• Name your Avatar and voice. "Keith with Keith's voice clone" beats letting it choose, the same way naming a person beats saying "a presenter" in any HeyGen prompt.

• Keep the folder lean. Cowork reads the folder before it works, and your usage is one shared weekly allowance across chat, Cowork and Claude Code. Five great past scripts beat five hundred random files.

• Check credits before a batch. If you're about to ask for several videos at once, ask it how many HeyGen credits you have left first. That's one of the built-in tools.

• Prefer a connector over the browser. Cowork can drive a browser, but it's slow because it works screen by screen. If the tool has a connector, use the connector.

• Your own tools need to be public to connect. Claude reaches custom connectors from Anthropic's servers, not your laptop, so a tool that only runs on your internal network won't connect. HeyGen's MCP is hosted, so this only matters for your own integrations.

The old version of this job was scrolling a week of Slack, deciding what mattered, writing it up, and getting on camera. The new version is three short replies and a face that was never in the room.

/$

7:44 PM · Sep 25, 2026

·

2,802

Views