YouTube2Text

How to Build an AI Agent with Claude Code (Claude AI Agent Tutorial) — Transcript

by AI Master · 3,776 words · 594 segments · language en · Watch on YouTube

Full transcript

  1. 0:00Most people are still using Claude like
  2. 0:02a fancy search engine. Ask a question,
  3. 0:04copy the answer, ask the next one.
  4. 0:06There's a way smarter way to work with
  5. 0:09it. And once you see it, you can't go
  6. 0:11back. Let me show you what an AI agent
  7. 0:13actually is and how to build one inside
  8. 0:16Claude without writing a single line of
  9. 0:19code. Before we build anything, I want
  10. 0:21to get clear on one thing because the
  11. 0:23term AI agent gets thrown around
  12. 0:25constantly and most explanations are
  13. 0:27either too technical or too vague to be
  14. 0:30useful. Here's how I think about it.
  15. 0:32There are three distinct levels of
  16. 0:34working with AI and most people are
  17. 0:36stuck at level one without realizing
  18. 0:39levels two and three even exist. Level
  19. 0:41one is basic chat. You ask something,
  20. 0:44Claude answers. You need a definition, a
  21. 0:46quick summary, an explanation. You type
  22. 0:49it in, you get it back. This is
  23. 0:50genuinely useful, but it's also the most
  24. 0:53limited possible way to work with the
  25. 0:55technology. You're essentially using a
  26. 0:57very sophisticated search engine. Level
  27. 0:59two is what I call builder mode. You
  28. 1:02stop asking questions and start asking
  29. 1:04Claude to make things. Write me this
  30. 1:06script, draft this email, analyze this
  31. 1:09document. Now you're getting real
  32. 1:11output, not just information. But the
  33. 1:13catch is that you're still managing
  34. 1:15every single step yourself. You decide
  35. 1:18what comes next, you review each output,
  36. 1:20you carry the whole project forward
  37. 1:22manually. Level three is agentic work.
  38. 1:25This is where the real shift happens.
  39. 1:27Instead of managing every step, you hand
  40. 1:30Claude a goal, a real complete outcome
  41. 1:33and it figures out how to get there. It
  42. 1:35thinks through the task, breaks it into
  43. 1:37phases, asks questions if something is
  44. 1:39unclear, executes each phase in
  45. 1:42sequence, reviews its own work, and
  46. 1:44delivers a finished result. You're not
  47. 1:46in the loop for every micro decision.
  48. 1:48You set the destination and the agent
  49. 1:50drives. And what makes Claude
  50. 1:52particularly powerful for this right now
  51. 1:54is that you don't need any developer
  52. 1:56setup to get started. The Claude desktop
  53. 1:58app already has a built-in code
  54. 2:00workspace where you can create projects,
  55. 2:03manage files, and run agentic workflows
  56. 2:05without touching a single line of actual
  57. 2:08code. That used to require serious
  58. 2:10technical infrastructure. Now, it's a
  59. 2:12desktop app. If you're running a
  60. 2:14business and juggling 10 different tools
  61. 2:16right now, what if I told you that one
  62. 2:18platform could replace all of them? This
  63. 2:20is GoHighLevel. You get access to pretty
  64. 2:23much all the features. Social media
  65. 2:25planner, you can schedule posts, see
  66. 2:27your content calendar, and even repost
  67. 2:29content across multiple platforms with
  68. 2:31one click. Email marketing, GoHighLevel
  69. 2:34has a full email suite. You can create
  70. 2:36campaigns, build email templates with a
  71. 2:39drag-and-drop editor. GoHighLevel has a
  72. 2:41built-in calendar and booking system.
  73. 2:43This replaces tools like Calendly. You
  74. 2:46can create different calendar types,
  75. 2:48one-on-one meetings, group calls,
  76. 2:50round-robin scheduling if you have a
  77. 2:52team. And here's the best part.
  78. 2:54GoHighLevel has a ton of pre-built
  79. 2:57workflow templates. I've linked 30-day
  80. 2:59free trial in the description, not the
  81. 3:01standard 14 days, an exclusive extended
  82. 3:04trial. Not every Claude conversation is
  83. 3:07actually agentic. There's a real
  84. 3:08difference between a sophisticated
  85. 3:10prompt and a genuine agent workflow, and
  86. 3:12it comes down to three things. The first
  87. 3:14thing is that a real agent follows a
  88. 3:16process, not just responds to a message.
  89. 3:19A chatbot gives you one reply per
  90. 3:21message. An agent moves through stages.
  91. 3:24It gathers context, forms a plan,
  92. 3:26executes the plan step by step, reviews
  93. 3:29its own output, and adjusts if something
  94. 3:31isn't working. It's operating through a
  95. 3:33system. The second thing is that a real
  96. 3:35agent makes decisions under uncertainty.
  97. 3:38A chatbot will guess if it doesn't
  98. 3:40understand your request. A real agent
  99. 3:42stops and asks. It adapts its approach
  100. 3:44based on what it finds. If the initial
  101. 3:46direction turns out to be wrong, it
  102. 3:48corrects course rather than pushing
  103. 3:50forward with a bad assumption. That
  104. 3:52ability to reason through ambiguity,
  105. 3:54rather than just pattern matching to the
  106. 3:56most probable answer, is what separates
  107. 3:59an agent from a fancy auto complete. The
  108. 4:01third thing is clarification before
  109. 4:03execution. This one is probably the most
  110. 4:06underrated. The vast majority of bad AI
  111. 4:08output happens because the model
  112. 4:10misunderstood what you actually wanted.
  113. 4:13It assumed an audience, a format, a
  114. 4:15scope, a tone. A well-designed agent
  115. 4:18asks before it acts. It treats your
  116. 4:20initial request as a starting point for
  117. 4:23a brief conversation, not as a complete
  118. 4:25specification to execute immediately.
  119. 4:27That clarification step is where most of
  120. 4:30the quality lives. All right, let's
  121. 4:32build something. Open the Claude desktop
  122. 4:34app and go to the code workspace. If you
  123. 4:37haven't downloaded it yet, it's
  124. 4:38available directly from Anthropic. Just
  125. 4:40search Claude desktop app and it'll be
  126. 4:42the first result. Once you're in the
  127. 4:44code workspace, create a new project
  128. 4:47folder. You can name it anything. I'll
  129. 4:48call mine agent workspace for this demo.
  130. 4:51This folder is your AI's operating
  131. 4:53environment. Think of it as the office
  132. 4:55your agent works in. Everything it
  133. 4:57creates, organizes, and references lives
  134. 4:59here. Right now, it's completely empty,
  135. 5:01and that's totally fine for what we're
  136. 5:03about to do. The first thing we need to
  137. 5:05add is the most important file in the
  138. 5:07entire setup. It's called Claude.md, and
  139. 5:10I genuinely believe most people who
  140. 5:12struggle with AI agents are struggling
  141. 5:15because they skip this file entirely.
  142. 5:17And creating it is genuinely the easiest
  143. 5:20part of this whole setup. You don't open
  144. 5:22a text editor. You don't mess with the
  145. 5:24file system. You just tell Claude to do
  146. 5:26it. Type create a file called Claude.md
  147. 5:30in the root of this project right into
  148. 5:32the chat, and it'll spin up the file for
  149. 5:34you in seconds. That's it. Now, we just
  150. 5:36need to fill it in. Here's the problem
  151. 5:38with most agent setups. People spend
  152. 5:40hours designing elaborate workflows,
  153. 5:43then launch them without giving Claude
  154. 5:45any context about who they are, what
  155. 5:47they're trying to accomplish, or how
  156. 5:48they want things done. And then they
  157. 5:50wonder why the output feels generic.
  158. 5:53Claude.md solves this. It's a plain text
  159. 5:56file written in markdown. Just a simple
  160. 5:58formatting system that Claude
  161. 6:00automatically reads whenever it starts
  162. 6:02working inside your project. Think of it
  163. 6:04as an onboarding document for your AI.
  164. 6:06Except instead of onboarding a new hire
  165. 6:09once, you write it once and Claude reads
  166. 6:11it every single session in your
  167. 6:12workspace. Let me walk you through what
  168. 6:15goes into a good Claude.md file because
  169. 6:18the structure matters. Start with a
  170. 6:20project context section. Two or three
  171. 6:23sentences about what this workspace is
  172. 6:25for. Something like, "This workspace is
  173. 6:27used for AI-assisted research, content
  174. 6:30creation, and workflow automation.
  175. 6:32[music] Outputs are primarily for
  176. 6:34non-technical audiences who want
  177. 6:36practical, actionable information about
  178. 6:38AI tools." Next, add an about me
  179. 6:41section. Tell Claude who you are and how
  180. 6:43you work. your background? Who's your
  181. 6:45audience? What tone do you prefer? For
  182. 6:47example, I create practical AI content
  183. 6:50for everyday users. I prefer concise,
  184. 6:53direct explanations over academic or
  185. 6:55jargon-heavy language. Every output
  186. 6:57should feel like advice from a
  187. 6:59knowledgeable friend, not a corporate
  188. 7:01report. Then write a rules section. This
  189. 7:03is where you encode your working
  190. 7:05preferences so Claude applies them
  191. 7:07automatically every time. Some rules I
  192. 7:09use in my own setup. Always ask at least
  193. 7:12three clarifying questions before
  194. 7:14starting any complex task. Always show
  195. 7:17your plan before you execute it. Keep
  196. 7:19every written output at or under the
  197. 7:21requested length. Never pad. When saving
  198. 7:24files, use lowercase names with hyphens,
  199. 7:27not spaces. These might sound like small
  200. 7:29things, but when they're encoded in the
  201. 7:31file, you stop having to re-explain them
  202. 7:33every session. Finally, define your
  203. 7:36folder structure. Something like a
  204. 7:38workflows folder for your instruction
  205. 7:40files, an output folder for completed
  206. 7:42work, and a resources folder for
  207. 7:44reference material you want Claude to
  208. 7:46draw from. This gives the agent a clear
  209. 7:49map of where things live and where to
  210. 7:51put things when it's done. Once you save
  211. 7:53this file, Claude's behavior inside your
  212. 7:56project changes immediately. It's no
  213. 7:58longer working from scratch every
  214. 8:00session. It has context, rules, and a
  215. 8:02structure to operate within. That one
  216. 8:05file, done well, is worth more than most
  217. 8:08of the prompt engineering tricks you'll
  218. 8:09find anywhere online. Before we create
  219. 8:12our first real workflow, I want to talk
  220. 8:13about one habit that makes or breaks AI
  221. 8:16agent work. It's called planning mode,
  222. 8:18and the idea is simple. You never let
  223. 8:20the agent execute immediately. Here's
  224. 8:23why this matters. When an AI agent
  225. 8:25starts operating autonomously, it
  226. 8:28commits to a direction. It creates
  227. 8:30files, structures content, makes
  228. 8:32assumptions, and if any of those initial
  229. 8:34assumptions are wrong, which happens
  230. 8:36more often than you'd expect, you've
  231. 8:38wasted time and you're now correcting a
  232. 8:40half-finished project instead of
  233. 8:41steering a clean one. Planning mode
  234. 8:44means you always ask Claude to show you
  235. 8:46the plan before it does anything. You
  236. 8:48want to see how it interpreted your
  237. 8:50goal, what steps it intends to take,
  238. 8:52what files it plans to create, and what
  239. 8:54questions it still has. That review
  240. 8:57takes maybe 2 minutes, but it routinely
  241. 8:59saves 10 minutes of cleanup on the back
  242. 9:01[music] end. The simplest way to build
  243. 9:03this into every workflow is to add one
  244. 9:06line in your
  245. 9:08section, "Always present a written plan
  246. 9:11and wait for approval before beginning
  247. 9:13any multi-step task." Once that's there,
  248. 9:16Claude defaults to showing you its
  249. 9:18thinking before it acts. You review, you
  250. 9:21adjust, you approve, then it executes.
  251. 9:24Quick pause before we start building. If
  252. 9:27what you just heard about agents has you
  253. 9:29thinking, "Okay, I actually want to get
  254. 9:31good at this," we built a place exactly
  255. 9:33for that. It's called AI Master, and
  256. 9:35there are two things inside that work
  257. 9:37together. First, the course. It's called
  258. 9:39Generative AI Essentials, and it's a
  259. 9:41hands-on encyclopedia of prompt
  260. 9:42engineering and automation. Modules on
  261. 9:44system prompts, multimodal work,
  262. 9:46marketing, data analysis, real
  263. 9:48workflows, not theory. Second, our
  264. 9:50studio. Once you've learned something in
  265. 9:51the course, you go practice it in the
  266. 9:53same platform on every major model.
  267. 9:55ChatGPT, Claude, Gemini, DeepSeek, Grok,
  268. 9:58plus Seedance and Kling for video and
  269. 10:00Nano Banana for images. One [music] tab,
  270. 10:01no juggling 10 subscriptions. Links in
  271. 10:04the description. And honestly, the
  272. 10:05annual plan is the smartest move, up to
  273. 10:0742% off the platform and every
  274. 10:10generation you run inside. Okay, back to
  275. 10:12building. Now, let's look at the
  276. 10:14underlying architecture of what we're
  277. 10:16building because understanding the
  278. 10:18structure makes you much better at
  279. 10:20designing your own workflows later. Most
  280. 10:23agentic systems have three components,
  281. 10:25and they work together in a specific
  282. 10:27way. The first component is the workflow
  283. 10:30file. This is a written document,
  284. 10:32usually a markdown file, that describes
  285. 10:35process the agent should follow for a
  286. 10:37specific type of task. It covers the
  287. 10:39goal, the steps, the rules, the expected
  288. 10:42output format, and what to do if
  289. 10:44something goes wrong. Think of it as a
  290. 10:46standard operating procedure written in
  291. 10:48plain English. You're not writing any
  292. 10:50code here, you're really just writing
  293. 10:52out a process in plain words. The second
  294. 10:54component is the agent itself, Claude.
  295. 10:57Claude reads your workflow file,
  296. 10:59understands the objective, and acts as
  297. 11:02the coordinator. It decides what to do
  298. 11:04at each step, what questions to ask,
  299. 11:06when to proceed, and when to pause for
  300. 11:08input. It functions like a project
  301. 11:11manager who's read the SOP and is now
  302. 11:14running the process. The third component
  303. 11:16is the tool set, the actual capabilities
  304. 11:18Claude can use inside the code
  305. 11:20workspace, reading and creating files,
  306. 11:23organizing folders, searching through
  307. 11:25documents, analyzing and editing text.
  308. 11:27For most beginner and intermediate
  309. 11:29workflows, these built-in capabilities
  310. 11:31are more than [music] enough. You don't
  311. 11:33need external integrations or APIs to
  312. 11:35build something genuinely useful. Here's
  313. 11:38the key insight, and this one runs
  314. 11:39counter to what most people expect. The
  315. 11:42most important component is the workflow
  316. 11:45file, not the tools. A thoughtfully
  317. 11:47designed workflow with basic tools will
  318. 11:50outperform a poorly designed workflow
  319. 11:52with sophisticated integrations every
  320. 11:54single time. The intelligence is in the
  321. 11:57instructions, not the technology stack.
  322. 11:59That's actually good news because
  323. 12:01instructions are something anyone can
  324. 12:03write. All right, let's build something
  325. 12:05real. We're going to create a research
  326. 12:08agent, a system where you give Claude a
  327. 12:10topic and it handles the entire research
  328. 12:12and report writing process from start to
  329. 12:15finish. Open a new conversation inside
  330. 12:17your Claude code workspace. Don't give
  331. 12:20it a topic yet. Instead, tell it what
  332. 12:22kind of system you want to build. Here's
  333. 12:24the kind of prompt that works well for
  334. 12:26this. I want to design a research
  335. 12:28workflow. When I give you a topic, you
  336. 12:31should first ask me clarifying questions
  337. 12:33about scope, audience, and desired
  338. 12:35depth. Then, form a written plan. Once I
  339. 12:38approve the plan, you research the topic
  340. 12:40thoroughly, organize your findings into
  341. 12:42clear sections, and save a structured
  342. 12:44report to the output folder. Before we
  343. 12:46build this, show me your plan for the
  344. 12:48workflow itself. Notice what we're doing
  345. 12:50here. We're not jumping straight to the
  346. 12:52task. We're designing the system that
  347. 12:54will handle the task. That extra step is
  348. 12:57what turns a one-time response into a
  349. 12:59repeatable workflow you can run over and
  350. 13:02over with different topics. Claude will
  351. 13:05come back with a proposed structure for
  352. 13:07the workflow. It might suggest things
  353. 13:09like a clarification phase with specific
  354. 13:12questions that will always ask, a
  355. 13:14research phase broken into subtopics, a
  356. 13:16synthesis step where it connects
  357. 13:18findings across sources, and a review
  358. 13:20step before saving the final file.
  359. 13:23That's a real agent workflow taking
  360. 13:25shape, built in plain English, requiring
  361. 13:28zero code. Once the structure looks
  362. 13:30right to you, tell Claude to go ahead
  363. 13:32and write the actual workflow file.
  364. 13:34It'll create something like research
  365. 13:36agent.md
  366. 13:38inside your workflows folder. When you
  367. 13:40open that file, you're reading the
  368. 13:42operating instructions for your agent,
  369. 13:44written in plain language, fully
  370. 13:46editable, and reusable every time you
  371. 13:49want a research report on a new topic.
  372. 13:51Now, let's put the workflow to work.
  373. 13:53Open a new session inside the same
  374. 13:55workspace and tell Claude you want to
  375. 13:57run the research workflow on a specific
  376. 13:59topic. I'll use the example I actually
  377. 14:01ran recently. I want a research report
  378. 14:04on the current state of AI agents in
  379. 14:062026. What's working in practice, what's
  380. 14:09over hyped, and where is the industry
  381. 14:11actually heading? If your workflow is
  382. 14:13set up correctly, Claude should not
  383. 14:15immediately start writing a report. The
  384. 14:17first thing it does is ask questions.
  385. 14:20Answer those questions specifically.
  386. 14:22Don't just say make it good. Something
  387. 14:24like audience is non-technical
  388. 14:26professionals who are curious about AI
  389. 14:28but don't have a dev background. Keep
  390. 14:30the tone conversational. Focus on tools
  391. 14:33and workflows people are actually using
  392. 14:35today, not theoretical frameworks.
  393. 14:37Target length around 1,500 words. Claude
  394. 14:41then forms a written plan. Here are the
  395. 14:43sections I'll cover. Here's the
  396. 14:44structure. Here's what I'll prioritize.
  397. 14:47You review it, maybe adjust one or two
  398. 14:49things, then you approve it, and then
  399. 14:50the agent runs. What happens next is
  400. 14:53genuinely different from a normal chat
  401. 14:55experience. Claude works through the
  402. 14:57phases of the workflow sequentially. It
  403. 14:59researches, synthesizes, writes,
  404. 15:02structures the output, and saves the
  405. 15:04report to your output folder as a
  406. 15:06completed file. You're not managing any
  407. 15:08of those steps. You approve the plan,
  408. 15:10and now it's executing. That moment when
  409. 15:13you watch the agent move through its
  410. 15:15workflow independently is when it stops
  411. 15:18feeling like a chatbot and starts
  412. 15:19feeling like something that actually
  413. 15:21works for you. Here's where a lot of
  414. 15:23people don't realize how powerful
  415. 15:25persistent project context actually is.
  416. 15:27Once your agent produces an output, you
  417. 15:30don't restart from zero to make changes.
  418. 15:32Claude already has full context of what
  419. 15:35it just created, so refinements are
  420. 15:37fast. Say the executive summary in your
  421. 15:39report runs a bit long. You just say,
  422. 15:42"Trim the executive summary. I want it
  423. 15:44down to three key points." Claude
  424. 15:46updates that section without touching
  425. 15:48the rest of the document. Or maybe you
  426. 15:49want to add something. "Add a short
  427. 15:51section at the end comparing the top
  428. 15:53three AI agent tools by use case."
  429. 15:56Claude goes back into the document,
  430. 15:58references what it already wrote, and
  431. 16:00adds the new section in a way that fits
  432. 16:02the existing structure and tone. This is
  433. 16:04the compounding effect of working inside
  434. 16:07a persistent workspace. Each session
  435. 16:09builds on the last. The context
  436. 16:12accumulates. Instead of starting over
  437. 16:14every time you need something new,
  438. 16:16you're refining and expanding a living
  439. 16:18body of work. That's a fundamentally
  440. 16:21different relationship with AI than the
  441. 16:23copy-paste loop most people are still
  442. 16:25stuck in. Okay, so the research agent
  443. 16:27was useful, but let's build something
  444. 16:29that genuinely saves you time every
  445. 16:32single week. We're making a repurposing
  446. 16:34agent. You feed it one long script, and
  447. 16:36it spits out three short scripts and a
  448. 16:39full social media pack, each one as a
  449. 16:41separate PDF sitting in a folder ready
  450. 16:43to publish. Here's what we're actually
  451. 16:45building. You drop your main script into
  452. 16:48an input folder. You give the agent one
  453. 16:50instruction. It reads the whole thing,
  454. 16:52picks the three strongest moments, and
  455. 16:54turns each one into its own 60-second
  456. 16:57short script with a fresh hook. Then it
  457. 16:59writes a social pack, an X thread, a
  458. 17:01LinkedIn post, an Instagram caption, and
  459. 17:04finally it converts every artifact into
  460. 17:06a clean PDF and drops them all into an
  461. 17:08output folder. Now, let's write the
  462. 17:10system prompt. Same approach as before.
  463. 17:13We describe the role, the inputs, the
  464. 17:15steps, and the output format. The role
  465. 17:17is a content repurposing specialist. The
  466. 17:20input is a single long-form script in
  467. 17:22the input folder. The steps are: read
  468. 17:25the script, [music] identify three
  469. 17:27standalone moments that work as shorts,
  470. 17:29write each one with its own hook and
  471. 17:31payoff, then draft platform-specific
  472. 17:33social posts that match the host's
  473. 17:35voice. The output is four markdown files
  474. 17:38converted to PDF, saved to the output
  475. 17:40folder. I'll run it live. One prompt and
  476. 17:43you'll see the files appear in the
  477. 17:45folder one after another. Three shorts
  478. 17:47PDFs, one social pack PDF, from a single
  479. 17:51script in under 2 minutes. I've spent a
  480. 17:53lot of time running these workflows and
  481. 17:55watching other people try to get started
  482. 17:57with them, and the failures almost
  483. 17:59always come from the same handful of
  484. 18:01mistakes. Let me walk you through them
  485. 18:03so you can skip the frustrating part.
  486. 18:05Mistake one is skipping Claude.md.
  487. 18:08Without that file, Claude has no
  488. 18:10persistent context about who you are,
  489. 18:12how you like to work, or what standards
  490. 18:14your outputs need to meet. Every session
  491. 18:17starts from zero. The quality is
  492. 18:19inconsistent and you spend more time
  493. 18:21correcting output than you save by using
  494. 18:23an agent in the first place. Write the
  495. 18:25file before you build anything else.
  496. 18:27Mistake two is using vague goals.
  497. 18:30Telling Claude to do some research on AI
  498. 18:32will produce something that feels like a
  499. 18:34Wikipedia article. Telling it to
  500. 18:35research how non-technical professionals
  501. 18:38in marketing are using AI agents for
  502. 18:40content production, focus on tools with
  503. 18:42free tiers, and target an audience
  504. 18:44that's never written a line of code,
  505. 18:46produces something you can actually use.
  506. 18:48The specificity of your goal determines
  507. 18:50the quality of the output. This is true
  508. 18:52at every level of AI work, but it
  509. 18:54matters most in agent workflows because
  510. 18:57the agent is running multiple steps
  511. 18:59based on that initial goal. Mistake
  512. 19:01three is skipping the plan review. It
  513. 19:03feels like an extra step and the
  514. 19:05temptation is to just let it run and fix
  515. 19:07things afterward, but agents that run
  516. 19:10without a reviewed plan can go
  517. 19:12significantly off track. Correcting a
  518. 19:14half-finished multi-stage workflow is
  519. 19:16much harder than adjusting a plan before
  520. 19:18execution starts. Two minutes of review
  521. 19:21up front saves 10 minutes of cleanup.
  522. 19:23Always approve the plan first. Mistake
  523. 19:26four is not requiring clarifying
  524. 19:28questions. If your workflow file doesn't
  525. 19:30explicitly tell Claude to ask questions
  526. 19:33before starting, it will assume, and
  527. 19:35those assumptions are the source of most
  528. 19:37mediocre AI output. Build the
  529. 19:39clarification step into every workflow
  530. 19:41you create. The agent should always
  531. 19:43understand your specific requirements
  532. 19:45before it starts working, not after.
  533. 19:47Mistake five is trying to build
  534. 19:49everything at once. People get excited
  535. 19:51about agents and want to build a
  536. 19:53research system, a content planner, a
  537. 19:55CRM workflow, an email responder, and an
  538. 19:58analytics dashboard all in the same
  539. 20:00week. The result is five half-working
  540. 20:02systems that none of them trust or use.
  541. 20:05Start with one workflow. Run it until
  542. 20:08it's reliable and you understand how it
  543. 20:10behaves. Then build the next one. Once
  544. 20:12you have one or two workflows running
  545. 20:15well, the natural question is, "What's
  546. 20:17next?" Here's a practical progression
  547. 20:19for how to expand your agent stack
  548. 20:22without overwhelming yourself. The
  549. 20:23progression that works best starts with
  550. 20:26simple self-contained research or
  551. 20:28writing tasks. Things with a clear input
  552. 20:31and a clear output format. A research
  553. 20:33report, a content outline, a document
  554. 20:36summary. These are low-stakes,
  555. 20:38high-feedback tasks where you can see
  556. 20:40immediately whether the agent performed
  557. 20:42well and where the cost of a bad output
  558. 20:45is minimal. From there, you can move to
  559. 20:47more iterative workflows. Ones where the
  560. 20:49agent produces a first draft and then
  561. 20:51refines it based on your feedback across
  562. 20:54multiple rounds. Script writing works
  563. 20:56well here. So does editing and
  564. 20:58improvement workflows, where you give
  565. 21:00Claude a rough draft and it applies
  566. 21:02specific criteria to improve it. The
  567. 21:04more advanced level is multi-workflow
  568. 21:07systems where one workflow triggers or
  569. 21:09informs another. A research agent
  570. 21:12produces a findings report and then a
  571. 21:13content planning agent reads that report
  572. 21:16and generates video ideas based on what
  573. 21:18the research identified. That's a real
  574. 21:21pipeline. It's not as complicated to
  575. 21:23build as it sounds. Once your individual
  576. 21:25workflows are solid, the thing I want
  577. 21:28you to take away from this progression
  578. 21:29is that agent capability isn't something
  579. 21:32you unlock all at once. You build it
  580. 21:34incrementally one reliable workflow at a
  581. 21:36time and the system gets more powerful
  582. 21:39as each component gets more refined.
  583. 21:41That's your first two working agents,
  584. 21:43one for research, one for repurposing.
  585. 21:45If you actually build even one of them
  586. 21:47this week, you're already ahead of 90%
  587. 21:50of people who just talk about agents.
  588. 21:52For a deeper dive on building agents
  589. 21:54from scratch, I've got another video
  590. 21:56linked up here. And if you want to
  591. 21:58actually learn this properly instead of
  592. 22:00just watching tutorials, that's what we
  593. 22:02built AI Master for. Links in the
  594. 22:04description.

About this transcript

This page contains the full transcript of How to Build an AI Agent with Claude Code (Claude AI Agent Tutorial) by AI Master, generated from the public captions YouTube serves with the video. The transcript has 3,776 words across 594 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.