YT YouTube2Text

YouTube2Text → Transcripts → Make Video From Claude and ChatGPT — Descript's MCP Explained

Make Video From Claude and ChatGPT — Descript's MCP Explained — Transcript

by Descript · 11,880 words · 1,808 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00the streamcaster,
0:02you've got to know
0:05what my code of a look your rate limit
0:08throttles will stop.
0:10When it's waiting with class, and you
0:13can't escape my
0:16APIs. [music]
0:18We're calling you.
0:20Authenticate
0:22every tool.
0:24APIs,
0:26connecting you.
0:28APIs [singing and music]
0:31connecting you, calling you, routing you
0:33to our tools.
0:38You get with class. You're posed to
0:40stop.
0:42Integrations abound where both that is
0:45the sub class rock. I'm in the flow.
0:50And you're locking me in.
0:53Halos go to froze. Authorize.
0:56With my endpoints aligned,
0:58and I can't stop my
1:02APIs. [music]
1:03We're calling you.
1:05Authenticate
1:07every tool.
1:09APIs [music]
1:11connecting you.
1:13APIs
1:30>> [music]
1:33>> I'm PCing you with MC [music]
1:36beats.
1:37When I'm passing the JSON and connecting
1:40the streamcaster, [music]
1:43you've got to know
1:45what my code of a look.
1:48Your rate limit [clears throat]
1:49throttles
1:50>> Um um um um I'm like the taste of like
1:53like non-chicken nuggets.
1:55>> [music]
1:59>> So, these nuggets are made from chicken,
2:02but they're made to emulate the taste of
2:03non-chicken nuggets.
2:04>> Dope.
2:05>> Oh, that's much better.
2:08>> Edit all the blather out of your videos
2:10because my time is very precious.
2:12>> Oh, that's fire.
2:13>> Make it less teal and more cerulean.
2:16>> Sure.
2:17>> Replace your background with something
2:19more fun.
2:20Cold void of outer space. [music]
2:23>> Let's boost that sound quality.
2:25>> Emulate the taste of non-chicken
2:26nuggets.
2:28>> [music]
2:28>> Emulate the taste of non-chicken
2:30nuggets.
2:30>> Dope.
2:32>> It needs more style. [music]
2:34It needs more clips, more gifts, more
2:36more.
2:40Well, I might have made it too gnarly.
2:43>> [music]
2:56>> Hello and welcome to our webinar today.
2:59Every time I watch that clip, I regret
3:01that I have not just eaten chicken
3:03nuggets. So, apologies in advance for
3:05you seeing that regret on my face for
3:06the following hour. I'm one of your
3:09co-hosts. I'm Sam. I'm joined by the
3:11illustrious, the wonderful Teddy from
3:13our QA engineering team. We are so
3:15excited to be here with you today.
3:16Teddy, how are you?
3:18>> I'm good. Now I'm thinking about chicken
3:20nuggets, too. Maybe that's what I'll
3:21have for lunch after this webinar.
3:23>> [laughter]
3:23>> Do all of us have chicken nuggets after
3:25this?
3:25>> I think I'll go for like the Impossible
3:27nuggets.
3:28>> Oh, okay. Yeah, you know, whatever kind
3:29of nuggets
3:30float your boat.
3:32We're here to talk about a really
3:33exciting topic today. We've gotten a
3:35bunch of interest and excitement from
3:38you, the community, about using Descript
3:40in different ways. We're here to talk
3:41about making video from within your AI
3:44tools, never leaving the comfort of your
3:46Claude or your ChatGPT sessions. And
3:49we've got an action-packed agenda today.
3:52If you followed us in our webinar series
3:54a couple months ago, uh Teddy uh with
3:57some other colleagues hosted a series of
3:59office hours where we gave you some
4:01practical demos on how to set up and
4:04connect Descript to uh Claude or to
4:07ChatGPT, and we proved that you could
4:10upload media and make simple edits from
4:11your AI tools. Today, we're hoping to
4:13take a a use case approach where we
4:15actually show you not just how to plug
4:17the thing in,
4:19but talk about some real-world scenarios
4:20where it makes sense to make videos from
4:23uh an AI tool, and more importantly,
4:26what are some uh some things that become
4:29unlocked once you start using Descript
4:32remotely and starting to, you know,
4:34encapsulate the logic of how you create
4:36certain videos with in the a sort of
4:38thing that AI tools are really good at,
4:40um that will let you scale and uh and
4:42really push the sort of output that you
4:43want to create. So, we've got an
4:45action-packed agenda today. Teddy,
4:46anything else you want to share before
4:47we we jump in?
4:49>> No, I'm just really excited to see if we
4:51can get everybody going on how awesome
4:53the API is and just how easy it is to
4:56use. You know, it really opens up
4:58Descript to everybody, and that's my
5:00favorite part about it.
5:03>> Well, without further ado, I say we jump
5:05in cuz we have so so much to cover. Uh
5:07we'll just pull up our slides real
5:09quick, and we'll give you a quick
5:11overview of where we're going to go
5:12today. So, just to just to touch um base
5:14with everybody, today's webinar is going
5:16to be for uh people who uh either
5:19already are using AI tools or they're
5:22interested and open to using them, and
5:24they have a ton of videos that they'd
5:25love to make, but they're not getting
5:26done for whatever reason. I think this
5:27is probably a really relatable pain.
5:30Um
5:30uh unless something changes, and
5:32something has changed, and we're going
5:33to talk about that today.
5:35Uh
5:36you'll leave today's session with uh
5:39knowing how to connect Descript into
5:40Claude or ChatGPT or your AI tool of
5:42choice in just a few minutes.
5:44The use case that we'll walk through
5:45today is turning a long form recording
5:47into ready-to-post clips, but the steps
5:50that we'll talk through, the workflow
5:51that we'll show you, will work with any
5:53number of use cases where you'd want to
5:54create videos that you'd want to
5:56replicate more a bunch of create a bunch
5:58of instances of that kind of content um
6:00over time. And then we'll show you a
6:02workflow that can how to take that
6:04creation workflow and turn it into a
6:06skill that you could repack package up
6:09and basically
6:10push one button or trigger an automation
6:12and have the same creation process
6:14happen again and again with different
6:16videos as they flow in.
6:19Uh so this is a high-level overview of
6:22our agenda. I'll talk really briefly
6:25about the API and MCP. We'll quickly go
6:28through a workflow overview. We'll talk
6:30through some setup steps Teddy will
6:31guide us through and then we'll jump
6:32into the demo.
6:34So really quickly I'll just talk for a
6:35minute about the API and the MCP. If you
6:38don't know what these acronyms mean then
6:40you might think I might be having like
6:42an alphabet stroke. I promise you these
6:43letters mean something. APIs and MCPs
6:46are just fancy words to describe tools
6:49that let Descript speak with other
6:52software in the case of APIs or AI tools
6:54in the case of MCPs. These are
6:56interfaces that allow say another
6:59program to ask Descript something or
7:01send Descript something and ask it to
7:03perform certain tasks. Today we're going
7:05to focus on the MCP or model context
7:08protocol which lets Descript be
7:10summonable within an AI chat.
7:14Teddy how was my description? Anything
7:15you'd want to add? Anything you'd want
7:16to share on top of that?
7:18>> No, that was perfect. On point. I think
7:21another way you could think about API or
7:23MCP and you know, you might wonder why
7:26does this matter to me? And the idea is
7:29whenever you have to edit a Descript
7:30video nowadays, you have to go to
7:32Descript to do everything, but now
7:34Descript is coming to you in your
7:36workspace, whether that's in Claude or
7:39ChatGPT, and able to work with also all
7:42those other tools that you have in your
7:44workspace. So, we're bringing Descript
7:46to you with the MCP.
7:49>> Yeah, that's a really good point. I know
7:50like this is the this is like the
7:52tagline I think.
7:55And I think it's really great with the
7:56points that you mentioned.
7:59>> Teddy, I think it kind of unlocks a lot.
8:01So, here's like a system diagram I think
8:03that is like maybe helps you think about
8:04this. If you already working in Claude
8:06or ChatGPT or any of your AI tools, this
8:09the Descript MCP acts as a bridge that
8:11lets
8:12you from these AI orchestrators, these
8:15AI tools, connect into Descript. And
8:18it's cool that you can kick off a task
8:20remotely. We'll talk about that in a
8:21second.
8:22And sort of like if you want to upload a
8:23video and wait for it to transcribe, you
8:25don't need to load Descript and upload a
8:27video. You can just ask Claude to do it
8:28and just come you know, go get a cup of
8:29coffee or work on something else.
8:31But all of the context that you have in
8:34Claude or in ChatGPT, other connectors
8:36that you have, your
8:39voice or sort of skills that you've
8:40created or other docs or things that
8:43you've built up in these AI tools, all
8:45that context lives there. And when you
8:47bring video creation superpowers from
8:49Descript into those tools, then you can
8:52start to use bits of your context when
8:54you want to create videos. And I've got
8:56some use cases listed I think on a
8:57coming slide that will will talk about
9:00that more. So, yeah, I I think just not
9:02to beat it a horse, but I think the MCP
9:04and the API from Descript really unlock
9:07two main usage modes I think of them.
9:10Access is the first one. This is letting
9:13you run Descript from anywhere you
9:14already work in AI tools,
9:16which can inherit your context, but also
9:18run in the background, right? Like if if
9:20there
9:22is a workflow task, you are like you
9:23want to upload a video and make a bunch
9:25of edits, you could do that all manually
9:26in Descript.
9:27But we'll show you some tools today that
9:29let you start that and then walk away
9:30while that processing happens in the
9:32background.
9:34And then automation is is other one. We
9:35won't get to that today, but in this is
9:37going to be a series of webinars and in
9:39the next episode what we'll take is the
9:41sort of package scale we create today,
9:43instructions on how to create a social
9:45clip or a series of social clips from a
9:46video, and then we'll build a simple
9:49automation to trigger those clips to get
9:51created every time a new recording lands
9:53in a folder, say. And that's those these
9:55are tools that both the API and MCP
9:57offer.
10:00So, as I mentioned, today we'll use the
10:02MCP focused in Claude. That's just the
10:05the environment that we'll demo in, but
10:06the same steps that we take in Claude
10:07will work with ChatGPT or other AI
10:09tools. We'll create a video that we'd
10:11ideally like to replicate. And this is
10:13like
10:14we're going to show you a process of
10:16refining and basically creating and
10:17editing videos from the MCP. Doesn't
10:19mean you couldn't do that in Dscript or
10:21another tool. You absolutely could. But
10:23we want to showcase how
10:24iterating and and actually editing video
10:26from your AI tool is simpler than you
10:28might think.
10:30And we're going to refine and create
10:33we're going to continue to refine this
10:35video that we're going to be working on
10:36these clips until we're really happy
10:38with the output. And then we'll take all
10:40of those edits that we worked on and ask
10:42our AI tool to wrap that up in a skill,
10:44basically describing the steps that we
10:46took so that in the future, the next
10:47clip that we want to make, we don't need
10:49to go through that whole process again.
10:50We're doing that work one time, wrapping
10:53it up in a skill, and then the next time
10:55we want to create clips we basically
10:56push a button or invoke this skill and
10:58it will follow the same instructions
10:59that we've refined.
11:01And this will work best for certain
11:05types of videos, certain types of
11:06workflows. If you're a creative agency
11:09and you create, you know, Sundance films
11:11and very bespoke one-off high-craft
11:12works of art, this is not probably the
11:15tool for you. There are some inherent
11:17limitations in the sorts of edits that
11:19you can do when not looking at Dscript
11:21or doing it from afar, but this sort of
11:23workflow works ideally when you have a
11:26series of consistent, highly structured
11:28videos that you I want to make. And and
11:31maybe you'd love to have these videos,
11:33but you've never quite had the time to
11:34make them. This is the perfect scenario
11:36where you'd want to invest in a little
11:38bit of time and like the machine to make
11:40the video once and then the button to
11:42push to make more and more instances
11:44over and over again. Some use cases top
11:46of mind for me in thinking about so like
11:49every time a sales engineer has a call,
11:51imagine if you could take that
11:52recording,
11:53just cut out all the cruft and small
11:55talk and then make a a sizzle reel of
11:57demos that were shown so that you could
11:58share with your customer. Or taking the
12:00first pass of a webinar rough cut,
12:02cutting out all the ums and ahs like I'm
12:04saying right now from this webinar,
12:05imagine we could automate that and have
12:07that automatically happen in the
12:08process. Or translating a finished video
12:11from uh one language into n languages.
12:14Uh turning a blog into a video. Social
12:16clips that we'll we'll talk about today.
12:17Teddy, are there any like favorite use
12:19cases you have that uh are top of mind
12:21for you for this sort of thing?
12:22>> You know, what I really love is
12:24especially these days Underlord is so
12:26smart with its media understanding. So I
12:29love using the MCP to basically take a
12:31bunch of footage and on some recurring
12:33cadence look at okay, here's all the
12:35footage I've dropped into Google Drive
12:38or into my Dropbox folder in the past
12:40week. You get this is the vibe of the
12:42video I want to make. Go make me a rough
12:45cut and then I'll re- give me some
12:47options and I say, oh okay, I like
12:49number two on the menu. Let's start
12:51there and that's when I actually jump in
12:53and start editing.
12:55And the reason why the MCP makes this
12:57more powerful than just using Descript
12:59in app is because the MCP enables me to
13:02pull from those other places like Google
13:04Drive, like Dropbox seamlessly and also
13:07for me to use Claude's brain and
13:09Claude's knowledge to chat and
13:11brainstorm and vibe a little bit better
13:13so that what ends up going into Descript
13:16already is kind of like the the like
13:19cream at the top of the can of coconut
13:21milk compared to just like the watery
13:23stuff at the bottom, you know?
13:25>> Yeah, always got to be hunting for that
13:26cream.
13:27>> [laughter]
13:29>> I saw a question asked and answered in
13:30the comments up on screen. MCP stands
13:32for model context protocol. Basically,
13:34think of it as a way that uh
13:38um an interface that lets AI agents talk
13:40with software or your tools.
13:43It's a standardized protocol that a lot
13:45of applications have built to let their
13:47tools work with AI agents.
13:49>> Yeah, and then also related to that,
13:52where I see a lot of questions in the
13:54chat around how much does this cost
13:55anyway, which of course is a very valid
13:58question anytime we're using AI tools.
14:00So, it's kind of split between the
14:03tokens you'll spend in whatever your LLM
14:06of choice is
14:07and Underlord AI credits. So, anything
14:10you do in Claude or ChatGPT will cost
14:14Claude tokens, but then when Claude
14:16sends something off to Underlord and
14:19stuff happens that Underlord does, then
14:21that'll cost you AI credits in Descript.
14:24But, that's all part of your normal
14:26Descript plan. We don't charge the MCP
14:29as a separate billing plan or you don't
14:31need to buy anything new or different or
14:33special to use the MCP. You just connect
14:36and as long as you've got AI credits,
14:38you can get going.
14:40>> Yeah, just to add to what Teddy's
14:41saying, if like tokens, credits, all
14:43those terms don't even make sense to you
14:44and you're not even sure what that
14:45means, with every Descript plan, you get
14:49a certain amount of spendable units that
14:51let you use Descript. With any Descript
14:54plan, even free plans, you are able to
14:57access the MCP
14:59and use Descript. So, you don't need to
15:00have a paid plan to connect things, uh
15:02but your ability to import media or use
15:05different editing features will spend
15:07those consumables, those media minutes,
15:09those AI credits. So, depending on how
15:11much you want to produce, uh the plan
15:12that you have will potentially be a gate
15:15to how much you can get done. Uh but as
15:17you change your plan, then you'll like
15:18increase your limits and what what not.
15:19And like uh the MCP also lets you
15:22facilitate stuff with Descript, but also
15:24you're still using your AI tool of
15:26choice. Like this is our MCP doesn't
15:28have any bearing on like whatever credit
15:30allocation you might have in Claude or
15:31Chat GPT say. You'll just be using that
15:33to talk to Descript like you might be
15:35using it to talk to other things.
15:37>> Yeah, exactly.
15:39>> Okay, well let's jump into our scenario.
15:41So imagine this, there are just never
15:44enough clips. Our CEO, this is a true
15:45story,
15:46Laura does a ton of interviews on
15:48podcasts. She has great conversations
15:50and it generates a ton of really
15:51thoughtful and valuable material and
15:53we've always wanted to but never been
15:56able to really meaningfully repurpose
15:58all of that content. And we've also been
15:59wanting to post a bunch more on our
16:01social channels to generate thought
16:03leadership content and drive brand
16:04awareness. This is like a true story
16:05from Descript, real life here. So the
16:08opportunity that we're going to tackle
16:09today is
16:10how can we come up with a standardized
16:12process to create simple structured
16:15clips from these interviews, from these
16:17podcasts with a repeatable look and feel
16:19that feels brand aligned
16:22and then make that into a process that
16:23we can quickly or effortlessly invoke in
16:26the future. So every interview she does,
16:28we push a button, blink, clips fall out
16:30of that and then we can use them on our
16:31social channels. In the next episode,
16:33not today, but next week what we'll do
16:35is we'll take the thing we build today
16:37and then create an automation so that
16:39you can sort of set and forget this
16:41thing and clips will continuously flow
16:42out of new recordings as they come in.
16:44But today we'll be building the
16:45machinery to actually create the clips.
16:48I wasn't sure if I saw a question come
16:49up on screen. Yes, this session Okay,
16:50cool. Yeah, just so that you know, this
16:52will be recorded. So don't feel
16:53frantically like you need to take
16:55screenshots of every screen and take
16:57furious notes. You'll be able to rewatch
16:58this later.
16:59>> Mhm.
17:02>> Okay,
17:03we that was the 30,000 foot view
17:05scenario or 10,000 meter scenario for my
17:07metric friends. The 3,000 Oh, I'm going
17:10to not stop I'm going to stop saying
17:11units. Here's the Here's the rough
17:14workflow we're going to follow today.
17:16Uh
17:16going to kick us off by showing us how
17:18to connect descript to Claude and chat
17:21GPT. Probably among the more common AI
17:23tools. And then I'm going to pick it up
17:25and talk us through how we create one
17:28finished in this case clip. We'll create
17:30one finished video. That's going to be
17:32the starting point where as we refine
17:34and get the video looking better,
17:35looking more like what we want, um that
17:37process of refinement will help us
17:39capture the instructions we use to make
17:41that video really nice uh with
17:43prompt-based edits. We'll talk more on
17:44that in a sec. Then we'll be able to
17:46test that skill and then turn it into it
17:48something that we can trigger later.
17:51Uh and if there aren't any other
17:52questions, Teddy, do you want to take us
17:54away to uh showing us how to connect
17:56descript?
17:57>> Let's take it away. Get the show on the
17:59road. Okay, so I'm going to start with
18:01Claude first and then chat GPT.
18:04He No particular ordering or preference.
18:06That's just how I had my tabs set up. Um
18:09So from Claude, and note with Claude
18:12these same steps work in the Claude
18:13desktop app, too. I'm just on web cuz
18:16that's what I like to use.
18:17So you're going to click here on the
18:20customize
18:21these icon.
18:23And you should see in the sidebar
18:25connectors.
18:26So this is where you can tell Claude
18:28what apps to talk to. And for example,
18:31right now you can see I have linear, my
18:34issue tracking system we use here at
18:35descript. I have notion connected. I
18:37have Google calendar.
18:39But if I want to add some more stuff, I
18:41can click add up here and I have an
18:43option for browse connectors and add
18:45custom connector.
18:47Right now, we're going to go to browse
18:49connectors.
18:51And from here you'll see there's this
18:52whole slate of other apps that Anthropic
18:54has approved because they work well with
18:57Claude and will be seamlessly integrated
19:00once you just plug them in.
19:02And so we can search descript from here.
19:06Okay, and click there, descript.
19:09And I'm going to click connect.
19:14And so what you'll see is Claude will
19:17kick me to this window to connect to
19:19Descript. This is basically telling in
19:22my Descript account, yes, it's okay to
19:25accept data from Claude and to connect
19:27with Claude and work with it.
19:29So I'm going to click allow.
19:32And then I should be redirected back to
19:34Claude.
19:36And the way that you can test how does
19:39this actually work is I can say make me
19:42a little video in Descript. Make me a
19:461-minute
19:48Descript video.
19:50And what I've got on the brain is I made
19:5310 lbs of potato salad last weekend for
19:56the holiday weekend. And as I was
19:58thinking about, hmm, I've got uh
20:00potatoes, which are carbs, and yogurt,
20:02which is a dairy, is that a cereal?
20:10>> [snorts]
20:13>> Uh let's just take a moment while all of
20:14our brains break considering such a
20:17philosophically horrifying question.
20:29>> So I'm just going to give it a little
20:31prompt. Right now, for the purposes of
20:32making sure it's connected, you don't
20:34even have to do anything complex.
20:36And there was a question in the chat
20:37earlier, too, about how do I make sure
20:39this doesn't eat up a bunch of credits?
20:41You can basically do that by choosing
20:43which model you're using.
20:44And just like how you can choose your
20:46Claude model, I could also, if I wanted
20:48to, include something here telling it
20:51which Underlord model to use, too. So if
20:53I wanted Underlord to like go to the max
20:55and say, "Underlord, use Fable 5." I
20:58could do that in my prompt. But I'm just
21:00going to use whatever the default is
21:01because I'm just trying to do a quick
21:03little connection test.
21:07And so as we see, Claude is
21:09contemplating.
21:13>> We had a question come up. I just want
21:14to speak to you, Teddy. Um if someone
21:16goes to the connectors view and doesn't
21:18see
21:20um
21:21Descript, uh
21:23do you have any ideas on what they
21:24should do? I mean, of course, if you've
21:25got any questions about Descript in
21:27general, blanket statement, you can
21:28always reach out to our support team.
21:30There are a ton of very talented and
21:31thoughtful people who know a bunch about
21:33Descript and can help you sort things
21:34out. But, are there any other scenarios,
21:36Teddy, that you could imagine where they
21:37wouldn't be able to see the connector in
21:38the marketplace?
21:40>> I think one thing to check would be that
21:42comes right off the top of my mind is
21:44are you on an enterprise or personal
21:47Claude plan? Because if you're on an
21:49enterprise plan or like a team plan,
21:51your team admin can might not might
21:54limit what connections you can make,
21:56reasonably so, you know, because they
21:58want to make sure we're using the right
22:00things things that And so, I would reach
22:02out to your enterprise admin about that
22:05as a first step. If you're on a personal
22:06plan though and you're still not seeing
22:08the connector, I would definitely reach
22:10out to our support team.
22:12>> Yeah, that's a great point. Um
22:14and you know, if if you're using
22:15Descript for work or you want to use
22:17Descript for work, uh and for whatever
22:19reason your IT admin's locked out any
22:21any like connectors, customer or
22:23otherwise, uh then it could be a good an
22:24interesting conversation to have,
22:26especially if you're already using
22:27Descript anyways, like, "Hey, this tool
22:28that we use for video, uh
22:30making videos gotten even easier. We can
22:32integrate it into our AI workflow. Could
22:33we like uh would you mind like could we
22:36explore adding this connector as a
22:37sanctioned connector in our enterprise?"
22:39>> Yeah. Exactly.
22:41And so, back to Claude. So, now we can
22:43see that Claude actually has contacted
22:46Descript. So, that's how we know the
22:48connection is successful.
22:50So, I'm just going to click always allow
22:51and let that run for now. Um but, we've
22:54basically successfully proved that
22:55Claude is connected to Descript, can
22:58talk to Descript, and if I I to Descript
23:01right now,
23:02I should probably be able to see the
23:04project that Claude has created over
23:06here, is potato salad a cereal?
23:10>> I want to um there's a question that I
23:12saw come up and I want to demystify some
23:13of the magic that Teddy just showed.
23:15Okay, so like you just asked a question
23:18>> Mhm.
23:18>> to make a video with Dscript.
23:21>> Yeah.
23:22>> How is that happening? Right? Like
23:24there's no media you're uploading. You
23:26are sending effectively uh
23:28vibes based sentence on what you want
23:30[laughter]
23:31and Dscript's doing it. Um the I guess
23:34the punchline here is Underlord is
23:35behind all of this stuff.
23:37Do you want to talk a little bit Teddy
23:39about like who Underlord is, what it
23:41does, and why it is the like the robot
23:44behind the curtain?
23:45>> Absolutely. So Underlord, as you said,
23:48is our robot behind the curtain who you
23:50can talk to in chat in Dscript anytime
23:52here. And basically, if you imagine that
23:55Claude and ChatGPT are LLMs with
23:59broad generalized knowledge of anything,
24:01Underlord is like like our personalized
24:05AI editor who knows a a lot of really
24:08good stuff specific to a video editing.
24:11So Underlord knows what makes a good
24:12video. It knows what when you should cut
24:16audio and how you should do it. When is
24:18a reasonable time to apply auto gain
24:20control. When might you want to apply
24:22studio sound because Underlord's a one
24:25job instead of knowing everything in all
24:28of humanity is to know everything about
24:30video and how to make the best video for
24:34you that it can. And so the magic here
24:37when we're in Claude and I say, "Hey
24:39Claude, I had make this video." Claude
24:42does a little bit of work first thinking
24:44about, "Okay, what do I want to tell
24:45Underlord to make?"
24:47But then the importing the stock media,
24:50the um writing, actually putting the
24:54script into the script pane, dividing it
24:56into scenes, you know, if I were to drop
24:59some audio or video tracks in here,
25:01picking what video should go where,
25:03that's all done by Underlord because
25:05Underlord has the specialized training
25:07and knowledge to know what what would
25:09make a good video and how what would
25:12make something that you would actually
25:13click publish.
25:16>> I also feel like it's safe to assume
25:17there are probably people out there who
25:18like are really AI savvy, they've used a
25:20bunch of different agents, they're
25:21familiar with agents built-in tools that
25:22like know more about their context in
25:24their domain. And one thing I wanted to
25:26talk about mechanically a little bit
25:27like how the API and MCP from Descript
25:30are set up. You can think of uh there
25:32being three primary tools.
25:43>> Oh, Sam, your audio is cutting out a
25:44little bit.
25:45>> Oh, no, is it better now, Bing Bing?
25:46>> Yeah, it's better now. Yeah.
25:47>> Okay, sorry.
25:48>> No, that's good.
25:49>> One of the tools you can use with the
25:51with the MCP is to import media, upload
25:53a recording, upload from a URL.
25:56Um the second tool that you can do is
25:58publish. If you've got a project already
26:00edited and you want to generate a share
26:01page or a download link, um you can use
26:04the MCP and API to do that. The third
26:06one and the most interesting one is the
26:08agent tool, the editing with prompt
26:10tool. We don't have a specific button on
26:13a on MCP button, a cloud command to
26:15apply studio sound and to trim this and
26:17to trim that. Instead, we push all of
26:19our edits through this agent API or this
26:22like a prompt-based project editing.
26:24>> Mhm.
26:24>> And that might seem limiting, but in
26:26fact, it's quite freeing. When all of
26:28the edits you apply are through the
26:29prompt, we support one avenue for making
26:32edits and that's basically loose
26:33instructions to Underlord who can then
26:35execute those edits on the back end. And
26:37the power of that is that Underlord,
26:39like Teddy mentioned, our AI co-editor
26:41that knows everything about video and
26:43everything about Descript, can take any
26:45and all of your instructions and then
26:46apply them as loose, as specific, as
26:49grand or as finite as they are to edit
26:53your video in turn. And that basically
26:55lets the sky be the limit in terms of
26:57the nature of editing commands you send
26:59in because Underlord can just handle it
27:01for you. And absolutely like anything
27:03with with AI, with anything
27:04prompt-based, like it takes it takes
27:06time to refine the prompt. And if you
27:08put a bad or a you know a
27:11a lame prompt in, it's like not well
27:12refined, you're going to get output
27:13that's reflective of that. But at the
27:15same time, um it's easy to refine and
27:17experiment and explore. And once you
27:19dial in a prompt, um and there's some
27:21other tools in Descript that you can
27:23lean on for consistency like layout
27:25packs, brand studio styles, colors that
27:29take the variability out of like
27:30Underlord understanding exactly what you
27:32mean. You've got enough tools to make a
27:34a wide swath of edits with enough
27:36consistency and controls to keep your
27:38sort of creation solution space in a
27:39finite place. Um
27:41I didn't mean to like go on a tangent
27:43there, but I wanted to kind of touch on
27:44like import, edit, and publish being the
27:46three main paint brushes that you have.
27:48>> Yes. And actually relatedly, while Sam
27:50was talking, you might have noticed this
27:52video is actively changing as we're
27:55watching. And like hands hands up here,
27:58I'm not doing anything to do that. That
27:59is all Underlord deciding like you might
28:02have noticed this B-roll it put in first
28:04was landscape. And Underlord said, "Huh,
28:07that doesn't make sense because we have
28:08a portrait
28:10shaped video right now. So I'm going to
28:11reshape it for portrait." And I didn't
28:13have to do anything because Underlord
28:16has the knowledge to do its first draft
28:19rather than "Hmm, okay, these are the
28:21things that I need to get back into
28:22shape so that what it serves to you at
28:25the end
28:26is pretty pretty darn good, if I do say
28:29so myself.
28:30But
28:31let's also I don't want to forget about
28:33ChatGPT because I know we have a lot of
28:35ChatGPT users as well. So let's show how
28:37to connect that real quick. It's pretty
28:39similar to the Claude flow,
28:41but just looks a little bit different.
28:43So here, logged into the
28:46you have instead this plugins
28:48button you can click over here on the
28:50sidebar.
28:52And from here,
28:55I should be able to search Descript as a
28:56plugin.
28:58Ta-da!
28:59And there it is.
29:01And I'll just click install plugin.
29:05And similarly with Claude, I'll need to
29:07sign in with Descript, basically telling
29:09Descript it's okay to work with ChatGPT.
29:11You know,
29:12this is something that I'm allowing and
29:14I've approved and it should take me back
29:16to ChatGPT.
29:22Okay, and now ChatGPT will tell me I can
29:24try this in chat.
29:28And it actually is This is fun is that
29:30ChatGPT already also will give me a
29:33basic prompt based on what people like
29:36to use Descript for. So, edit this
29:38recording for clarity, clean up my audio
29:40using some of our classic tools like
29:42Studio Sound or edit for clarity,
29:44removing filler words, and all that good
29:47stuff.
29:48And just because I know we've gotten
29:50some questions about this and
29:53I'll just show for example if I took
29:56some video that I have.
30:01So, here's for example
30:03a video I took at the aquarium of some
30:04jellyfish.
30:06I can say upload this into Descript,
30:09add soothing music, and captions
30:15about jellyfish and how great they are.
30:29Okay, and you can see ChatGPT has
30:31already contacted Descript. And so, it's
30:33going to ask me for permission. I'm
30:35going to grant it permission, but
30:36ChatGPT knows who Descript is,
30:38knows how to knows Descript's phone
30:41number, knows how to set up a calendar
30:43invite for Descript, etc., etc.
30:47And now we just let the let the bots do
30:49their thing and chug away.
30:52>> And the workflow from here basically
30:54seems quite similar, right? Like you can
30:55add prompts, you can upload media, you
30:57can edit, you can publish everything.
30:58The MCP interface with Claude and
31:01ChatGPT, there are some nuances that
31:02differentiate them, but for all intents
31:03and purposes, like this is the same.
31:05We're at the same point with both.
31:07Um and now the rest of this demo, we're
31:09going to jump into the clip creation now
31:11and um we're going to do a lot of
31:12kitchen swaps cuz I think what you'll
31:14realize here is like, you know, these AI
31:16tools take time. Uh
31:18also, all of this work executing work
31:20with Descript on the back end isn't
31:22instantaneous. It all takes time. But
31:23kind of the beauty of being able to
31:25trigger a lot of these actions in your
31:27AI tools is you can trigger them and
31:29then move on to literally do anything
31:31else uh and come back to them later. Get
31:34basically giving you background
31:35processing power with things like
31:37uploading media, waiting for it to
31:38transcribe. And when we get into more
31:40refined edits, actually doing the
31:41creation process all sort of on a
31:44background thread while you're doing
31:45something else.
31:46>> Yes. And then I just want to
31:48answer a few last questions I saw fly by
31:50in the chat real quick. One question
31:52that can I use the MCP on things I
31:54already have uploaded in Descript?
31:56Absolutely. You just connect Descript
31:59and Underlord and Claude have got your
32:00back. They know what's in Descript. You
32:04can tell them like, "Hey, I have this
32:05project from last week. I don't even
32:07remember the name, but I was working on
32:09it with these people and I think this is
32:10what it was about." And they will figure
32:12it out for you. Um and then you can
32:15even, I believe, pull in things from
32:17your media library as well, referencing
32:20it as you create your videos. So, if you
32:22have stock footage or, for example, an
32:24intro or outro you regularly use that's
32:26in your media library, you can reference
32:28that, too. You don't need to upload
32:30everything a la carte. You can also, if
32:33you have stuff waiting for you in
32:34Descript, just use that.
32:36>> Yeah, one of the my favorite ways to
32:38edit existing projects, I like to use
32:40Descript in the browser. So, I'll copy
32:41the URL of a project when I'm in a
32:43project, paste it into Claude and say,
32:46"Hey, I want to add captions to this
32:47project." Claude or ChatGPT, they're
32:49smart enough to strip the project ID
32:52from the URL, and you're off to the
32:53races. As Teddy mentioned, you can also
32:55search projects. And maybe one of the
32:57best ways to figure out all that you can
32:58do with the Descript MCP is ask your AI
33:01tool. Say, "Hey, list for me all the
33:02tools this MCP has." You can extract
33:05transcripts. You can list all of your
33:07projects. You can look for active MCP
33:09You can do a lot. And and your AI tool
33:11knows about all of those jobs.
33:13>> Yes.
33:15>> Okay, so with that, I think let's jump
33:16into the actual
33:18Oh, I think someone wants to
33:20Let's take a look at these jellyfish,
33:21Teddy.
33:22>> Let's take a look at the jellyfish. And
33:24this is one thing that's nice, too, is
33:26from Claude and ChatGPT, you can also
33:28open the project in Descript right
33:30there.
33:31Which I really enjoy. So, let's see. I
33:34think under Looks like it's still
33:36chugging away at this one, but it looks
33:38like my video has been uploaded.
33:41Let's see. Do I have to approve
33:42something? Ba ba ba ba ba ba. Okay,
33:44yeah. So, this one's still uploading,
33:46but when this finishes upload by the end
33:48of this webinar, I will publish and
33:50share the video just as an example of
33:53what our MCP can do.
33:54>> And if it's helpful, maybe I can start
33:55my demo with a finished product. You
33:57know, like those cooking shows where you
34:00see like the finished roast chicken with
34:02the beautifully glistening carrots and
34:03vegetables. Exactly.
34:05>> Exactly. Martha Stewart is like, "This
34:06is what we're going to make today. Now,
34:08on to chopping our carrots."
34:10>> Yeah, maybe we'll start that. Okay, so
34:11I'll switch the scene, and uh okay,
34:13again, usually I wouldn't do this. Uh I
34:15would show like the the process, but um
34:18I also see that there's a lot of
34:19interest about podcasts and like podcast
34:22content. Uh I am again, the demo I'm
34:25hoping to show you is I want to make
34:27social clips from podcast and webinar
34:29interviews.
34:31I We're going to work together to refine
34:33a prompt, a process to make those clips
34:36from these raw recordings, and I'll show
34:38you this is a finished product.
34:39Actually, let me Yeah, you won't hear
34:41the audio so there's not echo on my end,
34:43but I'll just show you what the finished
34:44product looks like so you can get a
34:46sense for
34:58>> Sam, your mic's cutting out a little bit
35:00again.
35:00>> Oh, no.
35:01>> Oh, now it's back.
35:02>> Okay, cool.
35:03>> Yeah.
35:04>> Um we can share these share pages later
35:06if you want to watch these videos.
35:07There's nothing special about them
35:09except they're you know, we've crafted
35:10them to be the clips that we want. This
35:12may not be the clip style that you want,
35:14but that is immaterial. What matters is
35:16we're going to go through the creation
35:18process of building a prompt that makes
35:20these sorts of clips from our raw
35:22recordings, and then you can follow
35:23similar steps on your end to recreate
35:25the sorts of uh
35:26clips or whatever sort of videos that
35:28you'd like.
35:29You'll notice there's B-roll, there's
35:30zooming in, there's captions, and then
35:33uh we've even got like Descript's logo
35:35from our brand studio interleave here.
35:38And the hope is that um
35:40this is a sort of thing uh even if you
35:42don't expect, you know, AI to be perfect
35:44at making this like 100% it's done, you
35:46don't need to touch it.
35:48Imagine if these clips were generated
35:49for me without needing to do anything at
35:5180% or 90% of the way there, and then
35:53instead of spending my time creating one
35:54from scratch, I could spend my time
35:57quickly refining and reviewing 10 of
35:59them and getting them all ready to to
36:01postable state. That's the workflow that
36:02we'll go through today. Okay, so that's
36:03the cooked chicken. Sorry there was no
36:05audio. Um
36:08so the whole process I'm going to show
36:09you here, I'm going to do a lot of like
36:10I said kitchen swaps where I'll show you
36:12the start of a prompt in Claude, and
36:13then we'll um back uh back out to like
36:15talk about the process of refinement,
36:17and Teddy's going to keep me honest if
36:18there are questions
36:19uh on the thread. This is a recording
36:21from a uh an interview our CEO did.
36:24And um the way to kick this process off
36:26is with Claude the same thing basically
36:28after Lara is on a podcast we'll often
36:30get the recording of the podcast from
36:32the host as a common courtesy. So I
36:34basically uploaded this media into
36:35Claude and this is the prompt that I
36:37gave it. Um I want to create one short
36:3960-second clip from this recording of
36:40Lara, format it in aspect ratio 1:1,
36:43find the moments where she's making
36:44provocative statements or assertions,
36:45the most daring and interesting points
36:47of the session, keep each clip to Lara's
36:49voice. I added this little this like
36:51suffix here to not using any existing
36:53skills cuz I'm also working on skills
36:55for this webinar. So I I wanted to like
36:57do it from scratch. You won't have to
36:58add that.
36:59Um and you're off to the races.
37:02Uh something you should know about
37:04Descript is that there are a ton of
37:06tools. I'll just open this up here. A
37:07ton of tools we already have for content
37:10repurposing. If you open the app, you'll
37:12see things like create clips, create
37:14highlight reels, find highlights. These
37:16are already pre-canned features in the
37:19app. And when you ask Claude to make
37:22these sorts of clips or or when you
37:25reference features similar to them,
37:26underlord and Claude are smart enough to
37:28try and guide you towards using some of
37:30these tools inherently.
37:32Um so the way that this process looks,
37:34I'll just show you how this thing
37:35unfolded. I know it's kind of boring to
37:36just like look at someone's Claude
37:37thing, so we'll spend a lot of time in
37:38Descript itself.
37:40But remember all of this could be
37:41happening um just from Claude without
37:44having the Descript app open. It
37:45uploaded the media that I dragged and
37:47dropped into Claude. Um it transcribed
37:50it and then it basically went and found
37:52the clip that I was looking for. This is
37:54the result of that first prompt. You can
37:55see it's a 1:1 1x1 aspect ratio and then
37:59the clip talks about um
38:02what is it? Oh yeah, uh
38:04ironically one-shotting um
38:06AI creation of videos uh
38:08being the sort of thing that makes
38:09creative content. The irony is uh is
38:11great. So this is like okay, it's a very
38:13simple clip. There's no captions. It's
38:16just a talking head. Um
38:18the process that we're going to go
38:19through here is basically using Claude
38:21to riff and refine giving it feedback,
38:24instructions, or requests on how we'd
38:25want to improve this video. Like one
38:27thing you might have noticed from the
38:28finished product is I'd love to add the
38:30have captions be shown. So, something I
38:32might ask Claude here is like let's
38:35captions and also use
38:48Now, I'm just like continuing this chat
38:49in the same thread. You can be as
38:51specific or as generic as you want.
38:54Generally, as you know, the more
38:56specific that you are with your prompts,
38:58the more specific of output that you'll
38:59get.
39:00But, every time we we add more feedback
39:04or like
39:05watch the video. I think like this is a
39:06recurring pattern that you'll see a lot
39:08in this like we are making one decent
39:10clip once so that we can take all these
39:12instructions later and bottle them. But,
39:14a really common workflow here is I'll
39:16watch this video
39:18the first prompt that I made. Maybe I
39:20like the clip that was picked. Maybe I
39:22think it cut off part of the the overall
39:24point that Laura was making. Maybe there
39:27are filler words that I wish removed.
39:29The process that we repeat over and over
39:31again is watching the result of what
39:34Underlord did. Oh, captions already
39:35coming in.
39:36Watching the result of what was edited.
39:38Taking notes on things we wish were
39:40different or improvements we'd like to
39:42make. Resubmitting them and then repeat,
39:45repeat, repeat. This is a simple albeit
39:47effective process that basically
39:50vibe edit your project in a direction
39:52>> Sam, your audio's
39:54got a blip again.
39:56>> Oh, man.
39:57>> Oh, you're back now. It seems like every
39:58time I I'm like, "Hey, Sam."
40:01Your audio immediately comes back
40:02without interference.
40:04>> Okay, maybe all right. Is it good now?
40:05You can hear me?
40:06>> Yeah, it's good. We can hear you.
40:08>> I'll just restate that last point.
40:10A really effective way to edit a project
40:13in this sort of fashion using the MCP or
40:15using Descript. But, the MCP is to um um
40:19suggest edits, like offer edits to
40:21Underlord, wait for those edits to
40:22finish, watch the resulting video, take
40:25notes on what you'd improve or changes
40:26you want to make, and then resubmit. And
40:28then, over and over again, watch,
40:30critique, watch, critique.
40:33And that might seem
40:34like indirect, right? You're like, why
40:37wouldn't I just edit with Underlord
40:39directly in Dscript? Or why would I like
40:41use the MCP? Well, what's
40:43what's nice about using the MCP is that
40:44you can easily background task um
40:47uh edits and commands if you're doing
40:49other work. But, what's more important
40:50and what we'll do in just a moment is
40:52all of the edits cumulatively that we
40:54make, like we just added captions, um I
40:56think music was at Yeah, you can see
40:59background music was added. We'll be
41:01able to take all of those edits
41:02afterwards in Claude and generate a
41:04skill that can be recalled from Claude.
41:07And that will help us trigger this over
41:09and over again.
41:10Yeah. Uh let's see.
41:11>> And also, to answer one of the questions
41:13that flew by on screen, around how do I
41:16come the file size limit if my file uh
41:19is a little too big for Claude or
41:20ChatGPT, that's also part of the beauty
41:23of using the MCP, actually, is because
41:25you can have it connected to Google
41:26Drive or Dropbox, which have much bigger
41:29file size limits. Or if you just have
41:31the media already in Dscript as well,
41:34the MCP can work with that. So, if
41:37you're if you're encountering that
41:39upload error of [snorts] Claude saying
41:41that this file is too big, never fear.
41:43There are many many paths around that.
41:47>> Yeah. Yeah, and like even if your file
41:48is hosted in some cloud servers, like
41:49Teddy mentioned, um the MCP and API can
41:52import files from an a URL as well. So,
41:54it's not like you need to always upload
41:56the file, like dragging and dropping it
41:57onto Claude. Wherever the file is, the
41:59URL is totally reasonable path to that.
42:01>> Yes. Yeah, that was just like the baby
42:03bones demo.
42:06>> So, uh I don't want to like chicken swap
42:08too many times here. But, as you can
42:11see, like a common workflow I like to
42:14follow is having Claude open in one
42:16window and Descript in another. I'll
42:18watch each videos
42:21edits after I've prompted them, take
42:23notes or critique them. I often like to
42:25just speak using like a dictation tool
42:27into Claude, and then resubmit them and
42:29refine, refine, refine until I'm happy
42:31with the output. And something that's
42:33really important and we'll talk about in
42:34a minute, you don't want to rush this
42:36step. Our goal is to create a repeatable
42:39process that makes good clips out of
42:42every future webinar. And we're just
42:43making this process manually once.
42:46It could be tempting to just say like,
42:48"Okay, this is fine enough. Let me like
42:49get to the automation part and just like
42:50see how it works." But, when you If you
42:53any process that you are going to build
42:55to replicate something, if what you're
42:56replicating is mediocre, then you're
42:58going to be replicating mediocrity.
43:00You're going to be really unhappy and
43:01always think like, "Oh man, I should
43:02have like really fixed the formatting
43:04and and achieved a better brand aligned
43:06style or something like that." So, you
43:08don't want to rush this process.
43:10Uh I have a couple different videos that
43:11show like different stages of this
43:13project where I added logos as an outro
43:16scene. I took like a single one-shot
43:18quote that describes the arc of the clip
43:21and put that as an intro scene. All the
43:23way up till the the finished revision,
43:26which again, all of this was just
43:27prompted with Claude asking more and
43:29more edits. This one, if I like show you
43:31these different scenes down here, you'll
43:33see there's an intro sequence with like
43:35the title. B-roll has been added at
43:38selective moments when Lara is saying
43:40something that's illustrative or has
43:42like some sort of visual component that
43:43could depict it. There's a caption style
43:46that feels legible, and uh and then I
43:48picked some music that I thought was
43:50nice. And the result was this set of
43:53clips. You can see we've got, you know,
43:54a bunch of different ones, all of which
43:56are on different topics, but each of
43:58which, let's see,
44:00Oh, these are the actual share pages.
44:01Each of which have this same style. It's
44:04as if like the same machine made each of
44:06these clips time and time again.
44:09Um there I don't know if there are any
44:10questions, but this would be a great
44:11time to ask specific questions around
44:13the process of refining the prompt uh
44:16through Claude cuz if there are any I'd
44:17love to touch them.
44:25And if there aren't, um
44:27one thing I'll step back on here is uh I
44:29wanted to jump into this project. So
44:31this was the actual Claude session I
44:32used to create those first few clips
44:36that I just showed you the finished
44:37product of. And you can see there's a
44:40lot of back and forth, right? Uh uh
44:43I asked for music, I asked for captions,
44:45um
44:46uh we would go back it would like make
44:47an edit that I wouldn't like and I would
44:49ask for it to refine it back and forth.
44:51Um
44:52when you get your video to a state where
44:54you're happy uh with the finished
44:56product, this is the part of the process
44:58where you can say, "Okay, I want to take
45:00all of the edits
45:02>> Oops, Sam, your audio cut out for a
45:03second.
45:05>> How about now?
45:06>> Welcome back.
45:07>> Uh it might be tab focus, I think if I
45:09leave that
45:10>> Oh, that's smart.
45:11>> You can uh once you're happy with the
45:13output, then the next step of this
45:16process is to take those edits and turn
45:18the instructions that got you from zero,
45:21like right from when you imported your
45:22media, all the way through to your
45:23finished output into what I'll call a
45:26skill. Skill is a common term used by
45:28these um AI tools to describe a
45:30repeatable set of instructions that you
45:32write once and apply many times. And
45:35thankfully, your AI tools are really
45:37good at creating these already.
45:39Because I did all of my refinements and
45:40edits in this thread, I basically said
45:43something like this, "Help me write a
45:44prompt that I can pass on to Oh, no,
45:47this is for the B-roll. I think the
45:48prompt that I wrote was like um
45:50basically, help me take all of the edits
45:52that we've made cumulatively and write a
45:56skill so that I could recall this in the
45:58future for for future clips, for future
46:00videos." And that
46:02basically what it will do is go through
46:04and write Here's an example.
46:07Um a full skill that you can then take
46:10and save as a skill in your Claude
46:12environment. And I'll show you how to do
46:14that right now.
46:15>> And to answer DIY Tourism's question on
46:17screen around will we provide commonly
46:19used prompts eventually, we actually do
46:22have a public repository of commonly
46:24used skills, like these skills that Sam
46:27is showing right now, that you can feel
46:29free to use. You can also contribute to
46:31if you're like, "Hey, this skill really
46:32works well for me, and I'd love to share
46:34it with other people as well."
46:37>> Yeah, we have a couple in there now, and
46:38we would love more contributions. I
46:40think what's tricky about these sorts of
46:43prompts is like the format of this clip
46:45might be very specific to the sort of
46:47thing we would want. And you might want
46:48to make different decisions. So, it's
46:50easy to share prompts, but it's hard to
46:52make a one-size-fits-all prompt that has
46:54like all the variability. Like, the
46:56beauty of it is these are so easy to
46:57create once you've got a video that is
46:59reflective of the finished product you
47:00want. You can make them incredibly
47:02personalized. In fact, I find the best
47:04skills are very um specific and and have
47:07like uh very, you know, specific
47:09instructions towards the end that you
47:11want because they're able to handle a
47:12lot of different inputs and get you
47:13close enough to the finished output.
47:15>> Mhm.
47:16>> Uh there's a question on screen, "Can
47:17you request specific changes? Remove
47:19B-roll footage of a guy drinking water
47:20in Claude, or do you have to do this in
47:22the Descript editor?" I think you should
47:23be able to do this in
47:24>> Mhm.
47:25You can do it You can do it wherever you
47:27want to. If you do it in Claude, you can
47:30do it there, and just know it'll cost
47:31you Claude tokens then, or you can do it
47:33directly in Underlord, and it'll cost
47:35you Descript AI credits instead. You
47:37know, it's really up to you based on
47:39which agent you feel like will better
47:42handle the request.
47:44>> Part of what makes this possible though
47:45is the media imported into Descript is
47:47processed, and Underlord basically, you
47:49know, watches the media and can
47:51understand things that are happening on
47:52screen. Uh so, because that happens,
47:55you're able to ask questions like, "Find
47:57the picture of the guy drinking water
47:58and remove that or something like that.
48:00And whether
48:02like Claude tries to do that on its own
48:04or go straight to Underlord, those
48:05visual descriptions exist along with the
48:07media and make it possible to do
48:09something like that.
48:11So quickly what I'll do just to finish
48:12this wing here. Here's the skill that
48:13was generated. After we made all of our
48:15edits in Claude, we refine, we added
48:17captions, we added music.
48:19Then I asked Claude to create a skill
48:21summarizing all of the instructions we
48:23made. I'll copy this and I'll go into my
48:26Claude menu into customize.
48:29And then under skills, I'll say add. I
48:32can create one with Claude. It will like
48:34ask and prompt you or I can write skill
48:35instructions myself. I'll just write
48:38myself and I'll basically paste those
48:40instructions into this field and then
48:42I'll call this like, you know, create
48:44clips. Whatever, whatever. I actually
48:46already did this with this exact same
48:48prompt. And this is it right here. It's
48:50exactly the output that you saw. It's
48:52got all of these instructions in here.
48:55And you can read through them. In fact,
48:56one thing I'd recommend is when you
48:57create the skill, what we'll do right
48:58now quickly, is test this skill on a new
49:01recording that it's never seen before.
49:03Evaluate the output and it's likely
49:05you'll want to make some changes or it
49:07doesn't perfectly recreate the sorts of
49:08things you liked from the first video.
49:10Well, the beauty is you can go back to
49:12that thread where you created the skill
49:14in the first place, just like I am here
49:16and say, "I tested this somewhere else
49:18and uh I didn't like this thing about
49:20it. Can we update the skill to make sure
49:22we always uh end on this kind of
49:24transition or add this sort of music
49:25instead?" These sorts of skills are the
49:27sort of thing you're going to want to
49:28refine, but as you refine them, what
49:30you'll find is you'll get closer and
49:31closer to the the desired output and the
49:33best way to do it is just to test it,
49:35see how the output looks, and then test
49:37again.
49:38Um yeah, let's jump into Q&A in just one
49:40second. The last thing I want to show
49:41you is once you have that skill, using
49:44it is really simple. In Claude, you say
49:46slash create
49:48uh this is the skill we wrote
49:51and I literally just
49:52uh call that skill, I'll upload a file
49:56from uh let's see, boink boink here's a
49:59new interview from Lara and I'll press
50:01go.
50:02And that's it. That is all of the
50:04instructions that we we crafted to
50:06create that skill are now being applied
50:07to a brand new recording. And this is
50:09the sort of thing in the next episode
50:11we'll use as a jumping off point to
50:12create an automation every time there's
50:14a new recording so you don't even have
50:15to do this. A new recording lands in a
50:17place and this automation can proceed.
50:20So we covered a lot. My throat is tired.
50:23I still
50:24but maybe it's a good time to jump into
50:27Q&A.
50:31>> Does it cost us credits in Descript to
50:33use the MCP? Yes, it does. It costs you
50:37credits just like using any of our AI
50:39tools cost you credits. So you'll see
50:42that build in your usage panel in the
50:44app that for example, if you edit a make
50:47an edit with an API, it should show up
50:50there that it cost me I don't know. I'm
50:52just just going to say 15 credits and
50:54it'll have a little tag next to it that
50:55says API as well so that you know what's
50:58coming from the API versus what's coming
51:00from you using it in app. But you do not
51:02have to buy a specific plan or any
51:06special kind of setup to use the MCP.
51:09>> Also yeah, using the MCP it doesn't cost
51:11extra credits because you're using the
51:12MCP. It's simply another way to create
51:15but it is not implicitly more or less
51:16expensive. The honestly the biggest
51:18factor in cost here, which we didn't
51:20really talk about, is the model that you
51:21use. I happen to use Opus 48 just by
51:24default. Uh
51:26experiment with models, right? Cheaper
51:27models uh lesser credit usage. It's the
51:29same as true with Claude, Descript, and
51:31any AI tool.
51:33>> Mhm.
51:38Does the model we use in Claude /
51:40ChatGPT be the same when we use the MCP
51:43for Descript? Um
51:45so if I'm interpreting your question,
51:48basically if you set where Sam was
51:51showing you could choose Opus or you
51:52could choose another model, that chooses
51:55the model that Claude uses,
51:57but that's not necessarily the same that
51:59Underlord uses. So, you could tell, for
52:02example, if you want Claude a combo, you
52:04could do is if you're like, "I don't
52:06need Claude to put that much brainpower
52:08into it." So, you could set Claude to
52:10say like Sonnet, but then you want to
52:12give Underlord a lot of brainpower. You
52:15could tell in tell Claude in your prompt
52:18that I want Underlord to use Opus. So,
52:20that way you're kind of redirecting your
52:23credit cost or like the brainpower of
52:26your agents to where it matters most to
52:28you.
52:29>> And if I'm not mistaken, uh we've
52:31recently expanded at least the API
52:32service if you're building custom
52:34applications, so that you can be more
52:35explicit in dictating which model is
52:36used. It was always there. I think it's
52:38a bit more explicit now. And um and I
52:40think the defaults when you use the MCP
52:42mirror the defaults that you would use
52:43when you use the app. I think it thinks
52:45like start defaults to Sonnet.
52:46>> Mhm, exactly.
52:50Can you explain more about what
52:51Underlord is? Underlord is our in-app AI
52:55co-editor. It's basically, imagine
52:57ChatGPT or Claude, but if ChatGPT or
53:01Claude was supercharged to know
53:04everything that's makes an amazing video
53:06and everything that is in Descript
53:09inside and out. So, that versus
53:11sometimes you might make a request to
53:13Claude or ChatGPT, and they go, "Mhm, I
53:15need that a little bit." Underlord
53:18already knows that because Underlord's
53:19number one job is to know about Descript
53:22and to know about what makes a good
53:23video.
53:26>> I feel like what it does more than
53:28anything is um
53:31do your bidding. It's like an eager
53:32intern who wants to like follow your
53:34instructions no matter what. Okay, I
53:35asked it to set the background to green.
53:38I don't know why it didn't do that.
53:41Um
53:42but basically, Underlord is the
53:44co-editing agent that you chat with
53:46Yeah, like Teddy mentioned, in the app.
53:48And all of the editing instructions
53:49you're giving Claude are just being
53:50passed along to Underlord, who has full
53:52context of tools, the project, the
53:54script. Um So yeah, many apps have an
53:58agent built into their tool. Underlord
54:01is Descript's.
54:02>> Yes. And then we have a question, is
54:04there a way to get Underlord to do a
54:06modified version of the shorten word gap
54:08feature that cuts pauses from middle of
54:10the gap so it doesn't risk clipping the
54:12end of the word. That's something that I
54:13would say give it a try yourself and let
54:15us know how it goes. But I will also
54:17offer we do have some
54:20just This is a little sneak preview is
54:22we do have some improvements to shorten
54:24word gaps coming hopefully your way
54:26pretty soon cuz I was working on those
54:27actually quite recently because we know
54:29it's a much beloved feature, but also
54:31that it's really frustrating when and
54:34the um the gaps get cut are not
54:36necessarily the gaps that you wanted to
54:37get cut.
54:39>> Teddy's response to me like often times
54:41uh
54:42when someone's like, "Hey, is it
54:43possible to do this with Underlord?"
54:45It's like maddening but also true.
54:47Usually the best answer is like
54:49probably. You should try.
54:51>> Give it a whirl, you know.
54:52>> Yeah, like these agents are so darn
54:53smart. Uh more often than not they do a
54:55halfway decent job if you like ask in
54:57specific terms what you want to try. So
54:58that's a really common theme and I
54:59encourage you all to try.
55:01Um okay, question. How do you save the
55:03workflow? So the workflow, if you're
55:05talking about the workflow uh Susan that
55:07we just talked through, making edits,
55:09refining refining, and then saying,
55:11"Okay, all of these edits, now I want to
55:12replicate them again." Um what I would
55:15do, and I'll just jump back to where I
55:17was here. Uh
55:19at the very end of all of the edits in
55:21the same thread where you made them, I
55:22would say, "Let's turn um
55:24My audio might be jumpy, so sorry in
55:26advance if it is.
55:38Any prompt that's spells anything like
55:40this, basically like take all of the
55:42edits that we
55:44and
55:44>> Can you focus again?
55:46>> Yeah, is it better now?
55:47>> Yeah, that's better. Thank you.
55:48>> Take all the edits we made, turn it into
55:49a skill so we can recall it again. Um
55:52any prompt that says basically like,
55:53"Hey, remember everything we did in this
55:55project? Skillify it for me." Um we'll
55:57basically ask we'll get Claude to spit
55:59out a big old prompt and then you cut
56:02and let's see this is what this is here.
56:05You copy that text
56:07and then in the skills menu of the
56:10customize section of Claude, you'll just
56:12say create a new skill
56:15paste those instructions and then call
56:17it whatever you want.
56:21Something like that. Create clips
56:22whatever. And then invoking that skill,
56:24using that skill
56:31is as simple as in chat
56:34going and saying
56:37create clips and then uploading your
56:39media.
56:43Oh, yeah.
56:44>> [laughter]
56:44>> Will we get to see jellyfish and potato
56:45salad? Well, the people hunger for
56:48jellyfish and potato salad.
56:50So, who am I but an agent of the people?
56:52Um
56:53here we can see first off the potato
56:55salad video and this one I want to note
56:57to that I even published this from
56:59within Claude because I looked at what
57:02Claude did and I said, "Okay, that looks
57:04good. Publish it."
57:06Two words, that's all I had to do and
57:08now I have a shareable descript link and
57:10I have a direct download to download the
57:13actual MP4 video file if I wanted to
57:16upload this to my Instagram, to my
57:17LinkedIn. You know, maybe I'm not going
57:19to upload upload this to LinkedIn, but
57:21you get the gist.
57:22Um and so I'll play it. I'm going to
57:24mute my mic though or I'm going to mute
57:26the audio though because I don't want
57:27there to be like horrible echoey sound
57:29bleed, but we can watch the video play
57:32for a few seconds.
57:37And so you can see this is like a real
57:38video that I might be served on my
57:40Instagram algorithm, for example.
57:43>> God, I love food analogies. This is
57:45killing me. Why are we looking at so
57:46much food during lunch?
57:47>> [laughter]
57:48>> Um and then if we look at the jellyfish
57:50one,
57:51this one I'm waiting on the publish to
57:53finish from ChatGPT, but we can see in
57:56the script the actual video that was
57:58made and you can see I've got I uploaded
58:01while Sam was talking. I uploaded a few
58:02more jellyfish videos because I realized
58:04my first one was only 8 seconds. So I
58:06just added some more.
58:07But you can see I've got some nice
58:09captions here like I requested about
58:11jellyfish. If I unfurl the timeline, I
58:14can also see some calming music has been
58:15added like I requested.
58:18So yeah.
58:19There you have it.
58:21>> Wow, we covered a lot. Let's quickly
58:23recap. There's some pro tips I'd love to
58:24throw up on screen, too. We connected
58:26Descript to Claude and ChatGPT. We used
58:30Claude to create a video, a clip from a
58:34raw recording. We refined that recording
58:36by giving iterative prompts, getting
58:37that clip closer to what we wanted. We
58:39asked Claude to create a skill for us
58:41describing all the edits beforehand and
58:44making it into a and then we made a
58:45skill so that we could recall that set
58:47of edits for future clips.
58:49Um some pro tips that made this work and
58:52like these are just things that I guess
58:53we've learned from experience. Vibing in
58:55a direction instead of trying to exactly
58:57replicate a certain type of video,
58:59generally is going to like you're going
59:00to be happy with the output. AI is great
59:02at being non-deterministic. Uh so if you
59:04want to exactly recreate something,
59:06there are tools to help you be more
59:07consistent, but it's generally just a a
59:09higher, you know, a higher barrier to
59:11get there.
59:12Um you can lean on Descript's brand and
59:14styling tools, though, to be consistent.
59:16We've got tools like layout packs and
59:17brand studio. Hopefully we can put some
59:19links in the chat to help you learn more
59:21about those things, but those let you
59:22define visual layouts and colors once.
59:26And then your prompt can refer to them,
59:28so that every video uses the same sorts
59:29of elements. Um, struggle with your art.
59:33We kind of glossed over that with our
59:34chicken swaps today, but like truly, if
59:36you want to automate something and you
59:38only have like mediocre results, you
59:41you're going to just replicate
59:42mediocrity. That's probably not what you
59:43want. So, struggle with making the first
59:46good version of whatever you want to
59:48make. Take your time. You'll be tempted
59:49to rush. Don't. And if you don't,
59:52dividends will be paid every time you
59:54use that skill you sweat into it to make
59:57um uh you know, high-polished output.
59:59And then we don't have time to talk
1:00:00about this, but there is like AI tool
1:00:01hygiene to help you like pull in your
1:00:03context. Another use case we can explore
1:00:05in the future is like pulling stuff that
1:00:06Claude has access to that isn't from
1:00:08Descript, like docs or spreadsheets or
1:00:11you know, call recordings and that sort
1:00:12of thing. Uh there is going to be
1:00:14another episode next week where we take
1:00:17the skill that we just developed
1:00:18together and we build uh think of it
1:00:21like baby's first automation. Uh a
1:00:24simple automated workflow anyone can
1:00:26build to make it so that every new
1:00:28recording that Lara, our CEO, uh is in
1:00:30will automatically create end clips like
1:00:32that. That's going to be our workflow,
1:00:33but you can map this anything that you
1:00:35want. And if you'd like to learn more or
1:00:37you'd like us to focus on different
1:00:39topics, please reach out to us. Let us
1:00:41know. This is a very exciting topic for
1:00:43us. You've also got some other learning
1:00:44opportunities on screen. Learn Descript,
1:00:46which is a set of modules inside the
1:00:48app. There's a weekly um community
1:00:50office hours where um really awesome
1:00:53people showcase the product and answer
1:00:54questions. There's weekly support
1:00:56sessions and then our next webinar is
1:00:57going to be uh next week.
1:00:59Any sign-off thoughts? Any uh closing
1:01:01jellyfish thoughts, Teddy?
1:01:03>> Uh any I think my closing thoughts are
1:01:06now after all this talk of chicken, I'll
1:01:07get some potato salad and roast chicken.
1:01:09I'm ready to go eat lunch. But this was
1:01:11a great webinar. I'm so happy to teach
1:01:13people. And I look forward to seeing you
1:01:15all next week.
1:01:16>> Thank you all so much. If you'd like to
1:01:18book a demo or talk with someone who can
1:01:19help you understand what a workflow with
1:01:20Descript looks like to you, there's a QR
1:01:22code on screen. Thank you all for tuning
1:01:24in. Have a great rest of your week, and
1:01:26may we all enjoy chicken nuggets.
1:01:28>> See you later.
1:01:47>> You know why you tried to send a mail
1:01:50from body? Cuz you tried to do it with
1:01:53the full one, you see. Token into
1:01:57Gemini.
1:01:59I still know I'm locked into my APIs.
1:02:03We're calling you. Authenticate [music]
1:02:07>> [singing]
1:02:07>> every tool.
1:02:10APIs.
1:02:11Connecting you.
1:02:13APIs.
1:02:18APIs.
1:02:18>> [music]
1:02:19>> We're calling you. Authenticate
1:02:23every tool.
1:02:25APIs.
1:02:27Connecting you.
1:02:29APIs. [singing]
1:02:33APIs.
1:02:34>> [music]
1:02:35>> We're calling you. Authenticate

About this transcript

This page contains the full transcript of Make Video From Claude and ChatGPT — Descript's MCP Explained by Descript, generated from the public captions YouTube serves with the video. The transcript has 11,880 words across 1,808 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.