YT YouTube2Text

YouTube2Text → Transcripts → Learn 80% of Claude Cowork in Under 20 Minutes

Learn 80% of Claude Cowork in Under 20 Minutes — Transcript

by Jeff Su · 3,501 words · 540 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00You're in the right place if you use
0:01Claude Chat everyday, are intimidated by
0:03Claude Code, and know Claude Co-work
0:05exists but not why you need it. Because
0:08in this video, we're breaking down the
0:09seven core capabilities of Claude
0:11Co-work, each with real-world use cases,
0:14so that by the end, you know exactly
0:16what it can do and where it fits in your
0:18workflow. Quick heads-up, I've been
0:20making AI videos for about 3 years now,
0:22and I've never felt this way about a
0:23single tool. Let's get started. Kicking
0:26things off with the three biggest
0:27differences between Claude Chat, the
0:29chatbot we're all very familiar with,
0:31and Claude Co-work, the native desktop
0:33app. First, while both Claude Chat and
0:35Co-work require an internet connection,
0:38Chat uploads files to the cloud, which
0:40limits you to 20 files per conversation
0:42and 30 megabytes per file. But, since
0:45Co-work can access your local files,
0:47neither of those limitations apply, and
0:49so Co-work can easily handle more files
0:51with larger file sizes. On top of that,
0:54Co-work has a much larger context
0:56window, meaning you can do a lot more
0:57work before triggering the compacting
1:00conversation function, which is
1:02basically Claude summarizing your
1:03previous context to free up space but
1:06potentially losing important details in
1:08the process. Difference number two,
1:10Claude Chat gives you a response in a
1:11chat window that you have to then copy,
1:13format, or download yourself. While
1:15Co-work actually does the work and
1:17delivers ready-to-use files directly in
1:19your folder. Spoiler alert, the final
1:21output isn't limited to just local files
1:24since Co-work can connect to external
1:29pages in Notion directly, for example.
1:32But, let's not get ahead of ourselves,
1:34we'll cover that later. Third, as a
1:35result of those first two differences,
1:37we need to prompt Co-work differently
1:39than Claude Chat. With Chat, we use
1:42task-first language, [music] so we tell
1:44it what to do step-by-step. For example,
1:46review my raw thumbnail photos and
1:48recommend a naming convention and folder
1:50structure I should use. You get back a
1:52recommendation in text, and then you go
1:54do the work yourself. With Co-work, we
1:56use outcome-first language, where we
1:57define the end result, the constraints,
1:59and the quality bar. For example, I have
2:0115 raw thumbnail photos in this folder.
2:03I need them organized into subfolders by
2:05topic with descriptive file names.
2:07[music] You let Co-work run, and 3
2:08minutes later, the work is done. If
2:10you're struggling with this, I prepared
2:12a prompt template where you can just
2:13describe what you're trying to do, and
2:15it spits out a prompt that's optimized
2:17for Claude Co-work, linked [music] down
2:19below. Now, let's go over a few
2:20essential settings so that Co-work is
2:22set up properly. First, make sure the
2:24Co-work tab is selected, and then under
2:27settings, the first general tab here,
2:30this is where you'll find personal
2:31preferences that apply to both Claude
2:34Chat and Co-work. But, if you go down to
2:37the Co-work tab, you'll be able to add
2:40instructions that apply to Co-work only.
2:43I've actually put together a starter
2:44pack of instructions beginners can just
2:46copy and paste immediately out of the
2:48goodness of my heart, obviously, and not
2:50at all because I'm trying to build a
2:51parasocial relationship so that one day
2:52you'll buy my paid products. What? No,
2:54no, that's just obviously a joke.
2:57Anyway, these instructions are basically
2:59guardrails that stop Co-work from
3:01touching your files without asking.
3:02Like, before deleting, overwriting, or
3:04renaming any existing file, show me what
3:07will change and wait for confirmation.
3:09Think of these like training wheels on a
3:10bicycle, right? You can remove them as
3:12you get more experience with Co-work.
3:14Next, under settings, capabilities,
3:16enable both memory features. I leave
3:18location metadata off, like that's going
3:20to help with data privacy.
3:22Um, set tool access to load tools when
3:23needed, and [music] enable all of these
3:26features down below. And finally,
3:28whether you're on Windows or Mac, head
3:30on over to your documents folder and
3:31create a new subfolder called Co-work
3:34Playground so that everything we do
3:36today is contained in that one folder
3:38and your other files stay untouched.
3:40Now, when we open up a new conversation,
3:42point Co-work at your playground folder
3:44by choosing the folder under documents,
3:46Co-work Playground, and if this is the
3:48first time you're giving access, it will
3:50ask you for access, and you can click
3:53always allow. And just know that Co-work
3:54is very strict about file access. So, if
3:57you drag a file from your downloads
3:59folder, for example, into Co-work,
4:01right? Co-work will be able to load it,
4:03but it can't read that file. So, if you
4:06want to follow along today, make sure
4:08all your files are within the playground
4:12folder. All right, capability number
4:13one. In plain English, Co-work can
4:15create, edit, and organize files
4:17directly on your computer. Diving right
4:20into an example, I've added a folder of
4:22over 100 receipts into the playground
4:24folder that are a mix of PDFs and JPEGs.
4:27>> [music]
4:28>> I tell Co-work, "Look, I need an expense
4:30report from the receipt photos in my
4:32receipts folder.
4:33Uh, give me Excel spreadsheet with a
4:35date, vendor, category, amount, and a
4:37totals row. If anything's blurry or
4:39unclear, mark it verify." Thanks to the
4:42starter pack instructions, Co-work first
4:44lists the steps it plans to take, waits
4:46for my confirmation, then reads every
4:48single image, extracts the relevant
4:50information, and outputs a formatted
4:52Excel file directly in my folder. And as
4:55you can see, there are also flagged rows
4:56for me to double-check. Now, is that
4:58possible in Claude Chat? Not really,
5:00because Chat has a 20-file limit per
5:03conversation, and we have more than 20
5:05receipts, obviously. And even if we had
5:07fewer receipts, Chat would still only
5:09give us a spreadsheet in a chat window.
5:11We'd then have to download and like save
5:13somewhere else ourselves. Example two, I
5:15have a massive 400-plus megabyte PDF
5:18that I need broken into smaller chunks
5:20so it's easier to upload onto AI tools.
5:22>> [music]
5:22>> I can just tell Co-work, "I have a
5:24massive
5:26PDF. I need it broken into separate
5:28files, one per chapter or major section
5:30with descriptive file names so I can
5:32find what I need at a glance." And
5:34Co-work, unlike Chat, will identify
5:38natural section breaks, create all those
5:41individual files,
5:43and then save them
5:44all back in your folder. Example three,
5:47Notebook LM can create amazing
5:49PowerPoint presentations, but the slides
5:51are images that can't be edited. So, we
5:53share the file with Co-work and say,
5:54"These slides aren't editable. Rebuild
5:56this as a clean, editable PowerPoint,
5:58same content, same slide order." And
6:01Co-work will first read all the slides
6:02to understand the underlying content
6:04before recreating the entire
6:05presentation with real text boxes you
6:08can actually modify.
6:10Although it's not perfect, as you can
6:12see here, it's better than just having
6:14static images. Persistent memory is the
6:16most important capability we're covering
6:18today, and it's only possible thanks to
6:20capability number one, Co-work having
6:22local file access. Let me explain with a
6:25silly example. If I go into Claude Chat,
6:27open personal preferences, and add,
6:29"Always address me as Heisenberg at the
6:31start of every conversation." I start a
6:32new chat, and Claude will always open
6:35with, "Hey, Heisenberg." Say my name.
6:38Heisenberg. You're goddamn right. That's
6:42technically persistent memory. You tell
6:44the AI something, and the AI remembers
6:46that thing in future conversations
6:49without you having to repeat yourself.
6:51But, because Claude Chat's memory is
6:53stored online, there's a hard limit to
6:55how much it can hold. Co-work saves
6:57memory to actual files on your computer,
6:58meaning it can remember every decision
7:00you've asked it to hold, every
7:01preference you've set for as long as you
7:04need. For example, if I ask Co-work,
7:06"How many newsletter editions have we
7:08produced together so far? Break it down
7:10by which app each edition covered." By
7:12the way, you should definitely sign up
7:12for my weekly newsletter if you haven't
7:14already.
7:15It'll load up the relevant context,
7:17[music] tell me we've produced seven
7:19editions together, two for Gmail, two
7:21for Chrome, and so on and so forth. And
7:23remember, this was from a brand new
7:24session with no prior conversation.
7:27Here's a use case you can try yourself.
7:28I have a meeting transcript within a
7:30subfolder in Co-work Playground, and I
7:32say, "Summarize the meeting transcript,
7:35limit to 200 words." And after it gives
7:38me that first output, let's say I don't
7:40like the structure of this, and I move
7:42the context up top, for example. Not
7:44that that's something I'd ever do
7:45because context always comes after, but
7:48I tell Co-work, "I've made some changes
7:50to your summary. Compare your version
7:52with mine and save those preferences in
7:53a visible root Claude MD and memory MD
7:56files so you remember them next time."
7:59Co-work then reconciles the two
8:01versions, figures out what I changed,
8:04and most importantly, and I can't stress
8:06this enough,
8:07most importantly, it creates something
8:09called a Claude MD and memory MD file.
8:13First, you can see that Co-work created
8:14these two files under your playground
8:16root folder. How these work is more
8:19advanced than what we'll cover today, so
8:20I'll save that for the next video. For
8:22now, just know whenever you tell Co-work
8:25to remember something, it writes to
8:27those files. And the more it writes, the
8:30better Co-work gets at working the way
8:33you want it to. So, keep those two files
8:35there. By the way, I recently revamped
8:36my free AI toolkit from scratch. New the
8:39original, this is a completely different
8:40experience, link down below. Next up,
8:43connecting Co-work to tools. By default,
8:45Co-work can only see what's in your
8:47folder. Connectors let it reach into the
8:49tools you already use, like Gmail,
8:51Google Drive, or Notion, right? So, it
8:53can read from them and work directly
8:55inside them. To set this up, click the
8:57customize tab on the left, go to
8:59connectors, click the plus icon, then
9:01browse connectors, and you can search
9:03for the tool that you use. At the very
9:05minimum, to follow along today, I
9:06recommend you connect Gmail, Google
9:08Calendar, uh, Google Drive, and Notion
9:10if you use it. Once you've connected
9:12Gmail, you can tell Co-work something
9:13like, "I want you to understand my tone
9:15of voice. Go through my emails from the
9:16past month to extract my tone of voice,
9:19and save those as writing style
9:20principles you can follow going
9:22forward."
9:25Co-work then proceeds to read through my
9:28emails, locates those with substantive
9:30writing, distills the patterns it sees,
9:33and presents them to me, and then saves
9:35them as writing style principles within
9:37the memory MD file so that next time I
9:40when I draft an email, it remembers my
9:42tone of voice. Next example, although my
9:44team and I take meeting notes in Notion,
9:46we use Google Gemini to auto-generate
9:50transcripts that are stored in Google
9:52Drive. So, if I were an anal,
9:53high-maintenance, unreasonable manager,
9:55which I'm not,
9:56I tell Cowork, "Hey, check the Gemini
9:59transcript against the meeting notes in
10:01Notion and surface commitments that
10:03didn't make it into the notes." And
10:04Cowork will read both documents, compare
10:07what was said versus what was written,
10:10and surface the action items, the
10:12commitments that didn't make it into the
10:15notes page. This is an example of two
10:17connectors working together via Cowork.
10:20It pulls a transcript from Google Drive,
10:22pulls the meeting notes from Notion,
10:24cross-references them, and surfaces what
10:26fell through the cracks. By the way, if
10:28you can't find the tool you want under
10:31connectors, you can add a custom
10:33connector through something called an
10:35MCP, which we're not covering today, but
10:37just know that option exists. All right,
10:39capability number four, skills. Let's
10:41dive right into a simple example. I just
10:43asked Cowork to write something for me,
10:45and I want to clean up the output. So, I
10:46run it through a prompt template I use
10:48all the time that makes any text more
10:50clear and concise.
10:52Cowork gives me a version two of the
10:54output. I like it, and now I can tell
10:56Cowork, "Hey, turn what you just did
10:59into a clear and concise skill." You can
11:02see that Cowork first runs its own skill
11:04creator skill for this task. Then, it
11:06asks me a few clarifying questions.
11:09Uh this skill should apply to any text.
11:11It should rewrite and provide a change
11:13log so I can see what changed. And yes,
11:16let's skip the test cases, and Cowork
11:19proceeds to keep going. After receiving
11:21permission, Cowork creates the skill,
11:24and I can now just add the skill to my
11:27Cowork with just one click.
11:29And clicking the manage pop-up on the
11:31top right lets me see all the skills I
11:33currently have access to. Now, the next
11:36time Claude gives me something and I
11:37say, "Make it more clear and concise,"
11:39Cowork simply runs its last output
11:41through the skill we just created. So,
11:44that was just a simple one-step skill,
11:46which honestly isn't that impressive,
11:47right? But the underlying concept is
11:50exactly the same for a weekly report
11:52that might take 10 steps from start to
11:54finish. You teach Cowork what to do, how
11:56many steps there are, and it repeats it
11:58every single time, which is a massive
12:00time-saver over the long run. Here's a
12:02more practical example. When I was at
12:04Google, different teams would send me
12:05their weekly updates in completely
12:07different formats, and I had to combine
12:08everything into one clean update for
12:11leadership. So, I share the raw updates
12:13from the three teams with Cowork and
12:14say, "Combine these three team updates
12:16into one formatted weekly marketing
12:18update I can send to leadership." After
12:21a few minutes, Cowork produces a first
12:23draft. And although it looks great, it's
12:25way too dense. So, I give feedback,
12:27"Lead with three top-line metrics across
12:30all teams, then give me three highlights
12:32and three lowlights. Keep it within 300
12:34words."
12:35Cowork revises, but gives me the output
12:37in chat. So, just for fun, I ask for a
12:40PDF format. And after a few more
12:42minutes, Cowork gives me the report in
12:44PDF, and the structure is exactly what I
12:47wanted. So, I can finally tell Cowork,
12:49"Now, go back through our conversation
12:51and create a weekly report skill that
12:53captures this entire workflow." After
12:55asking me a few clarifying questions,
12:57Cowork summarizes my request, asks for
13:00permissions again, then spends a few
13:02minutes creating the skill and running
13:04tests with hypothetical scenarios.
13:06Once that's all done, Cowork tells me I
13:08need to install the skill manually,
13:11almost like I planned for this to
13:12happen. So, I head on over to customize,
13:15skills, plus icon, upload a skill, find
13:18the skill.md file that Cowork generated,
13:20and boom, we now have a weekly report
13:23skill. Three more things you need to
13:24know about skills. First, to create your
13:26own skill, head on over to customize,
13:29skills, and make sure the skill creator
13:33by Anthropic skill is enabled. And while
13:35there are websites that offer skill
13:39templates, I highly recommend creating
13:41the first few yourself to get a hang of
13:44creating skills from scratch. Second,
13:46you can update a skill anytime by
13:47telling Cowork which skill to change,
13:50what to change about that skill, and
13:51granting permission for the change. But
13:54after it recreates it, you need to click
13:57copy again to overwrite the previous
13:59version.
14:00I'd also recommend backing up your
14:02skills to Google Drive because if you
14:03move to a new computer, skills do not
14:06transfer over. Third, under the
14:07customize option skills tab, you'll see
14:10if you click plus, there's an option to
14:12create a skill with Claude. I do not
14:15recommend doing this because it's much
14:17more effective for you to actually just
14:18go through the actual workflow first,
14:21go through that back-and-forth process,
14:23then reverse engineer the skill at the
14:25end, like I showed you earlier today.
14:27Moving on, at the risk of massively
14:29oversimplifying, Cowork projects and
14:31Claude chat projects are basically the
14:33same thing, except Cowork projects come
14:35with all the capabilities we've already
14:36covered. Local file access, persistent
14:38memory, connectors, skills, all of it.
14:41But there is one thing worth calling
14:42out. Cowork projects can write to their
14:44knowledge files directly, unlike Claude
14:47chat projects. Here's what I mean. I
14:49have a Claude chat project called
14:50Clarity Partner that helps me make my
14:52writing more articulate. After a session
14:54where we go back and forth on a piece of
14:55writing, Claude learns something new
14:57about my style, and I want to save that
14:59as a permanent principle so it remembers
15:02next time. In a chat project, after I
15:04ask Claude to codify our learnings,
15:06Claude is able to generate an updated
15:08knowledge file for me. And while I am
15:10able to add this to the project as a
15:12knowledge file in one click, I still
15:14have to manually delete the old file
15:17from the project settings. In Cowork, I
15:19can just say something like, "Codify
15:20this principle under the Clarity Partner
15:22project." Like when I'm explaining a
15:24concept, always start with an example
15:26first. And Cowork writes it directly to
15:28the instruction file without me having
15:30to do anything. If you're watching this
15:32and thinking, "I want a system like this
15:34for my own work," I'm actually building
15:35a course that walks you through exactly
15:37that step by step. I'll leave a link to
15:39the waitlist down below. So, [snorts]
15:40ironically, this next capability also
15:42highlights one of Cowork's biggest
15:44weaknesses. The capability itself is
15:46simple. If you have the Claude extension
15:48installed on your browser, Cowork can
15:50theoretically hand off tasks to the
15:52browser extension. I say theoretically
15:55because I simply can't recommend the
15:57extension right now for three practical
15:59reasons. First, it's slow as
16:01since most interactions require a
16:03screenshot sent back to Cowork before it
16:05decides what to do next. Second, it's
16:07unreliable. There were many times where
16:10it just stopped halfway through a task
16:12without finishing. Talk about being
16:14blue-balled by Claude. Third, it burns
16:16through your usage because it overthinks
16:18every step since it has to be super
16:19careful when controlling your browser.
16:22Coming back to the weakness I mentioned
16:23earlier, while both Claude chat and
16:25Cowork can search the web, Claude chat
16:28gives us a lot more control. We can
16:30force web search, enable or disable, and
16:33we can even enable or disable this
16:34mid-conversation. With Cowork, you can't
16:37force a web search, and it often falls
16:39back on the browser extension. And given
16:41how unreliable the extension is right
16:43now, that's a real limitation. To be
16:45clear, scheduled tasks aren't new.
16:47Gemini and ChatGPT have had this feature
16:49for a while now, but I found that only
16:51Claude Cowork's scheduled tasks work
16:53flawlessly, and it's because of the
16:55capabilities we've just covered. Diving
16:57into an example, let's open up scheduled
16:59tasks and click into this morning inbox
17:02triage task that runs every day at 6:00
17:05a.m. It produces a report and draft
17:07replies for all my emails. Let's
17:10actually just open up an example to show
17:12you.
17:13And because of how I set this up, the
17:17report and the draft replies are
17:18basically flawless. So, let me walk you
17:20through how this works. First, long-time
17:22viewers know I teach my inbox zero
17:24workflow in the Workspace Academy
17:26course, and Cowork took that inbox zero
17:28workflow and distilled a set of rules to
17:30follow. And thanks to capability number
17:33one, local file access, it saved those
17:36rules
17:37in a markdown document, as you can see
17:40here. Inbox zero triage workflow, and
17:42these are the instructions. Second,
17:45since Cowork is connected to Gmail
17:46through the connectors feature,
17:48capability number three, it's able to
17:50read
17:51all my emails, right? Map them against
17:55my inbox zero rules and what it knows
17:57about me to draft contextually relevant
18:01replies. But this did not work so well
18:03at the beginning. For the first week or
18:05so, I had to give feedback to Cowork on
18:07how I would have rephrased certain
18:09emails. And thanks to persistent memory,
18:11capability number two, Cowork remembers
18:14those corrections and now triages my
18:16inbox the way I would. And this does
18:18require some setup. In my actual
18:20Cowork folder, I have an email HQ
18:22subfolder, and here you can see within
18:25the Claude.md file, I have my inbox zero
18:28triage workflow, right? And within the
18:30memory.md file right over here, it
18:33remembers the feedback I've given it.
18:35For example, how I like my how I like to
18:37sign off in my email signature, right?
18:39So, this represents weeks of feedback
18:42and context and basically learning, so
18:44to speak. I know that looked
18:45intimidating, but don't worry, I'm going
18:47to dive much deeper in future videos.
18:49So, see you soon, and in the meantime,
18:52have a great one.

About this transcript

This page contains the full transcript of Learn 80% of Claude Cowork in Under 20 Minutes by Jeff Su, generated from the public captions YouTube serves with the video. The transcript has 3,501 words across 540 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 translate the transcript into 100+ languages.

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, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.