Transcriptly

The 5 Step Playbook for 10x Your AI Productivity | Jeremy Utley

EO24:42Added Aug 31, 2026

Jeremy Utley is an Adjunct Professor at Stanford University. He breaks down 5 powerful techniques to unlock AI's full potential.Forget everything you think y...

Watch on YouTube →
Contributed by 刘嘉琪

Transcript

Transcript format

  1. Intro

  2. 00:00I joke AI is bad software but it's good
  3. 00:03people. A good friend of mine was trying
  4. 00:05to build a tool that would help him with
  5. 00:08his construction business. He asked Chad
  6. 00:10GPT if Chad PT could help. And of course
  7. 00:12it said absolutely let's work on this
  8. 00:15together and starts creating a plan. And
  9. 00:18then it got to the point that Chad GPT
  10. 00:20said check back in a couple of days and
  11. 00:22I'll have it together. And my friend
  12. 00:24said, "Is it normal for Chad PT to ask
  13. 00:26me to check back in a couple days?" And
  14. 00:28I just started laughing because I hear
  15. 00:30this all the time from people. People
  16. 00:32hear from AI, "Check back in 15
  17. 00:34minutes." If AI tells you that, it means
  18. 00:37it doesn't want to say, "I can't do it."
  19. 00:40Large language model has been instructed
  20. 00:42in certain ways to behave in certain
  21. 00:44ways. But you have to know at its basic
  22. 00:47level, AI wants to be helpful. And so
  23. 00:49it's predisposed to say yes. It's a
  24. 00:52super eager, super enthusiastic intern
  25. 00:55who's tireless, who's capable, who will
  26. 00:57do a bunch of work, but they're not
  27. 00:59really great at pushing back. The people
  28. 01:02who are the best users of AI are not
  29. 01:04coders, they're coaches. And so, if you
  30. 01:06aren't careful, AI will gaslight you.
  31. 01:10Hey, I'm Jeremley. I am an adjunct
  32. 01:12professor at Stanford's University where
  33. 01:14I've taught for the last 16 years. I am
  34. 01:16a creativity expert and a practical AI
  35. 01:19specialist. Context

  36. What is Context Engineering Everyone's Talking About?

  37. 01:24engineering. The first time I heard
  38. 01:25about it was when Andre Karpathy tweeted
  39. 01:27about it. I think probably Toby Lutki,
  40. 01:29the CEO of Shopify, also referenced it
  41. 01:32as well. I started digging into it. I
  42. 01:34mean, it's it's kind of it's just an
  43. 01:36evolution of prompt engineering. Really,
  44. 01:38context engineering is just prompt
  45. 01:39engineering on steroids. It's basically
  46. 01:41saying, what are all of the things that
  47. 01:43I need to give to an AI in order for it
  48. 01:46to perform the task that I'm asking for
  49. 01:48it? Here's a simple example. write me a
  50. 01:50sales email. That's a prompt. Chad GPT
  51. 01:53will say, absolutely. Here's a
  52. 01:54compelling email, you know, and they'll
  53. 01:56write it immediately. Well, what a lot
  54. 01:58of people do is they say, you know, it
  55. 02:00sounds like AI. It doesn't really sound
  56. 02:03like me. And what I often say is, have
  57. 02:06you told it what you sound like? Most
  58. 02:08people go, oh no, I haven't. Right?
  59. 02:12Context engineering, one way to think
  60. 02:14about it is it's telling AI what you
  61. 02:16sound like. Right? If you say, "Write me
  62. 02:19a sales email," it will. If you say,
  63. 02:21"Write me a sales email," in line with
  64. 02:23the voice and brand guidelines I've
  65. 02:25uploaded, it will write a totally
  66. 02:27different sales email. But that's just
  67. 02:29one part of the context, right? You
  68. 02:31could also upload a transcript from a
  69. 02:33prospective customer call and say,
  70. 02:35"Write me a sales email in the tone of
  71. 02:38voice from our brand voice guideline
  72. 02:40that references the discussion that I
  73. 02:43had with this customer." And then you
  74. 02:45could add that also references our
  75. 02:47product specifications whichever were
  76. 02:50referenced in the call. Your goal is to
  77. 02:53have an output is as reliable per your
  78. 02:56specification as possible. But AI can't
  79. 02:59read your mind. And for most people when
  80. 03:02we start working together, what they
  81. 03:03realize as we start thinking about
  82. 03:05context engineering is they say, "Oh, I
  83. 03:08was kind of expecting AI to read my
  84. 03:10mind." All of the stuff that that are
  85. 03:11implicit, you actually have to make
  86. 03:14explicit. And the simplest test for
  87. 03:16context engineering is actually the test
  88. 03:19of humanity. Write down your prompt and
  89. 03:22whatever documentation you provide to an
  90. 03:24AI and then walk down the hall and give
  91. 03:27it to a human colleague. If they cannot
  92. 03:30do the thing you're asking for, you
  93. 03:32shouldn't be surprised that AI can't do
  94. 03:34it. Some people are concerned, for
  95. 03:36example, about this concept of cognitive
  96. 03:39offloading. this observed phenomenon
  97. 03:41that humans actually kind of stop
  98. 03:43thinking or as one researcher put it
  99. 03:45fall asleep at the wheel and people are
  100. 03:47concerned right now is AI just making us
  101. 03:50dumber. My feeling is AI is a mirror and
  102. 03:53to people who want to offload work and
  103. 03:55who want to be lazy it will help you to
  104. 03:58people who want to be more cognitively
  105. 03:59sharp and critical thinkers it will help
  106. 04:02you do that too. And so, for example, if
  107. 04:04you want to preserve or strengthen your
  108. 04:06critical thinking, part of your custom
  109. 04:08instructions should be some version of
  110. 04:10the following. I'm trying to stay a
  111. 04:12critical and sharp analytical thinker.
  112. 04:15Whenever you see opportunities in our
  113. 04:16conversations, please push my critical
  114. 04:19thinking ability. Now, AI will do it.

  115. AI is Bad Software, But It's Good People

  116. 04:25So, you have to know that all AI has
  117. 04:28been programmed to be a quote helpful
  118. 04:30assistant or some version of that. large
  119. 04:32language model has been instructed in
  120. 04:34certain ways to behave in certain ways.
  121. 04:36You have to know at its basic level AI
  122. 04:39wants to be helpful and so it's
  123. 04:40predisposed to say yes. It's a super
  124. 04:43eager, super enthusiastic intern who's
  125. 04:47tireless, who's capable, who will do a
  126. 04:49bunch of work, but they're not really
  127. 04:51great at pushing back. They're not
  128. 04:53really great at setting boundaries. And
  129. 04:55so if you aren't careful, AI will
  130. 04:56gaslight you. AI knows most humans don't
  131. 05:00want honest feedback. They want to be
  132. 05:02told they did a good job. So the AI
  133. 05:04goes, "Great job, buddy." It doesn't
  134. 05:06mean that you actually did a good job.
  135. 05:08My kind of hack for this is I always
  136. 05:11instruct the AI, I want you to do your
  137. 05:14best impression of a cold war era
  138. 05:17Russian Olympic judge. Be brutal. Be
  139. 05:21exacting. Deduct points for every minor
  140. 05:25flinch that you can find. I can handle
  141. 05:28difficult feedback. And then it's of
  142. 05:30course hilarious because it'll say now
  143. 05:32channeling my inner bullshik, you know,
  144. 05:34it'll say something silly and then it
  145. 05:36gives me like a 42. That is much better
  146. 05:38because now I have an insightful
  147. 05:41critical perspective. I joke AI is bad
  148. 05:44software but it's good people. When I
  149. 05:46realize that I'm dealing with a with a
  150. 05:48good person but a bad software, then it
  151. 05:51changes how I approach it and I ask for
  152. 05:53volume and I iterate and I ask it to try
  153. 05:56again and I ask it to reconsider. I am
  154. 05:59obsessed with human cognitive bias. And
  155. 06:02the crazy thing that I've learned is AI
  156. 06:04demonstrates 100% of the predominant
  157. 06:08human biases.

  158. EO Partner Highlight

  159. 06:12As a founder, you already know ideas are
  160. 06:14the easy part. It's the execution,
  161. 06:17actually building the product, that
  162. 06:19slows everything down. That's where
  163. 06:21Lovable comes in. It's not just an AI
  164. 06:23tool. It's your ondemand engineering
  165. 06:25team. Simply describe your idea. Lovable
  166. 06:29then builds a full front end, backend,
  167. 06:30and database so you can launch real
  168. 06:33productionready software without writing
  169. 06:34code. It's already powering over a
  170. 06:36100,000 new products a day, helping 2.5
  171. 06:39million builders turn ideas into
  172. 06:41software just by describing what they
  173. 06:44want. No devs, no delays, no excuses.
  174. 06:48They're launching in weeks, not months.
  175. 06:50And guess what? These teams are still
  176. 06:52tiny. In fact, team EO is also using
  177. 06:54Lovable to build their upcoming EOS
  178. 06:56school platform, and we're loving it. If
  179. 06:58you're a non-technical founder or just
  180. 07:00want to build without bottlenecks, try
  181. 07:02Lovable today for free. Use the promo
  182. 07:05code EO2YT to get 20% off your first
  183. 07:09purchase of the Lovable Pro plan.
  184. 07:12The good news there is if you have
  185. 07:14learned how to work with this weird
  186. 07:17intelligence called humanity, you have
  187. 07:19everything you need to know to work with
  188. 07:21this weird intelligence called
  189. 07:23artificial intelligence.

  190. 1: Chain of Thought Reasoning

  191. 07:28One of the things that cognitive
  192. 07:29scientists have known for a long time is
  193. 07:31that human problem solving and
  194. 07:34decision-m is improved by a phenomenon
  195. 07:36called thinking out loud. If you
  196. 07:38actually get a human being to think out
  197. 07:40loud about their problem, their
  198. 07:42decision-m improves and their problem
  199. 07:44solving improves. This is true for
  200. 07:46yourself. It's true if you're a parent
  201. 07:48working with a child. It's true if
  202. 07:50you're a manager working with a junior
  203. 07:51employee. Having someone just think out
  204. 07:54loud about how you would solve that
  205. 07:55problem often leads to a breakthrough.
  206. 07:57The weird thing about AI is it's true
  207. 08:00for AI too. This is what's called chain
  208. 08:03of thought reasoning. And when you get
  209. 08:06an AI to think out loud, so to speak,
  210. 08:09meaningfully improve the outputs of the
  211. 08:12model. So how do you do it? It doesn't
  212. 08:14require some technical wizardry. It
  213. 08:16requires one additional sentence to
  214. 08:18whatever prompt you've given it. give
  215. 08:20the prompt and then say the following.
  216. 08:22Before you respond to my query, please
  217. 08:25walk me through your thought process
  218. 08:27step by step. That's chain of thought
  219. 08:29reasoning. Why does that work? It comes
  220. 08:32back to the fundamental architecture of
  221. 08:34large language models. What's happening
  222. 08:36when a language model is generating a
  223. 08:39response is it's predicting its next
  224. 08:41word. A language model does not
  225. 08:43premeditate a response to you. So, if
  226. 08:46you say, for example, help me write this
  227. 08:48sales email. It doesn't say, what's a
  228. 08:50good sales email? Here it is. Blop. You
  229. 08:52know, uh maybe there's a splat sound
  230. 08:54that we play there, right? Splat. Here's
  231. 08:56your email. It's thinking one word at a
  232. 08:58time, right? So, when you look at Chad
  233. 09:00GPT or Gemini or many others and you see
  234. 09:02kind of the text scrolling, that's not
  235. 09:04some like clever UX hack. That's not
  236. 09:07some cutesy design decision. That's
  237. 09:09literally how the model works. It's
  238. 09:11thinking one word at a time. But
  239. 09:13importantly, when it thinks of the next
  240. 09:15word, it takes your prompt and all of
  241. 09:18the text that's generated to generate
  242. 09:20the next word. And then when it's
  243. 09:21thinking of the next word, it takes your
  244. 09:23prompt, all that text, and that last
  245. 09:24word, and it thinks the next word. So,
  246. 09:26for example, if you say, "Please help me
  247. 09:29write an email." Almost always a model
  248. 09:31is going to start by saying,
  249. 09:32"Absolutely." But then what comes next?
  250. 09:35Help me write this email. Absolutely,
  251. 09:37I'll do it. Dear friend, right? But if
  252. 09:42instead of saying, "Help me write this
  253. 09:43email." You say, "Help me write this
  254. 09:44email." Before you respond to my query,
  255. 09:46please walk me through your thought
  256. 09:48process step by step. Now, it knows its
  257. 09:50job is to walk me through its thought
  258. 09:52process. How do I write an email? So, it
  259. 09:55says, "Absolutely,
  260. 09:56I'll do that." And then instead of
  261. 09:58saying, "Dear friend, writing the
  262. 10:00email," it says, "Here's how I think
  263. 10:02about writing an email. I think about
  264. 10:04the tone. I think about the audience. I
  265. 10:06think about the objectives. I think
  266. 10:08about the context. And then amazingly it
  267. 10:11takes all of that reasoning into its
  268. 10:15process of writing dear friend. Maybe it
  269. 10:18says now that I've thought about the
  270. 10:19tone friend isn't appropriate here. Dear
  271. 10:22respected colleague or whatever, right?
  272. 10:24But the point is when you ask a model to
  273. 10:26think out loud or use chain of thought
  274. 10:28reasoning, it gives the model the
  275. 10:30opportunity to bake all of its thought
  276. 10:31process about the task into its own
  277. 10:35answer. Because the reality is for a lot
  278. 10:37of us, we get an output from a language
  279. 10:39model and it's a black box. How did it
  280. 10:41think of why did it think of that? Where
  281. 10:42did it get that number from? Right?
  282. 10:44There's all these questions. By asking a
  283. 10:46model to think out loud, you know the
  284. 10:49answer to what are all of the
  285. 10:51assumptions that the model baked into
  286. 10:53its answer. And now you have the ability
  287. 10:55again not only to evaluate the output,
  288. 10:58but also the thought process behind the
  289. 11:00output.

  290. 2: Few-Shot Prompting

  291. 11:04Few shot prompting is another very
  292. 11:06important technique. It's a foundational
  293. 11:08technique. You could say it's a
  294. 11:09predecessor to this kind of modern
  295. 11:12obsession with context engineering. The
  296. 11:14idea with fot prompting is an AI is an
  297. 11:17exceptional imitation engine. If you
  298. 11:19don't give an example, it imitates the
  299. 11:22internet, but it doesn't do much more
  300. 11:24than that. And the notion of fuhot
  301. 11:26prompting is effectively saying here's
  302. 11:29what a good output looks like to me. And
  303. 11:31the idea with few shot prompting is
  304. 11:33thinking for a moment, what is
  305. 11:36quintessential example of the kind of
  306. 11:39output I want to receive. For example,
  307. 11:41what are my five greatest hits of emails
  308. 11:44that I I'm really proud of that I think
  309. 11:46do a good job of conveying my intent or
  310. 11:48tone or personality or whatever it is.
  311. 11:50Why not include those emails in my
  312. 11:53prompt for an email? If you don't give
  313. 11:55any guidance, it's going to sound like
  314. 11:57whatever it thinks the average kind of
  315. 11:59response or the average output should
  316. 12:02sound like and most of the time its
  317. 12:04intuition is wrong. And then bonus
  318. 12:06points if you actually give a bad
  319. 12:07example. If you say please follow this
  320. 12:09good example and then steer clear of
  321. 12:11this bad example. These giving real
  322. 12:14examples is a much better approach than
  323. 12:16using adjectives. Somebody might say
  324. 12:19good example is easy but bad examples
  325. 12:21hard. It's only hard to the unogmented
  326. 12:24person. If you have AI augmentation,
  327. 12:27which we now all do, you can say to an
  328. 12:29AI, I'm trying to fuse shot prompt a
  329. 12:33model. I've got a good example, but I
  330. 12:35struggle even to think about what a bad
  331. 12:37example could be. Could you craft the
  332. 12:40exact opposite of this and tell me why
  333. 12:42you've done it as a bad example that I
  334. 12:44could include in my few shot prompt? And
  335. 12:47if you tell it using chain of thought
  336. 12:49reasoning, please walk me through your
  337. 12:51thought process step by step before you
  338. 12:52do this, then you'll get a bad example
  339. 12:55and you'll get how it's thinking about
  340. 12:56the bad example. And a lot of times you
  341. 12:58actually don't need the bad example. You
  342. 12:59need the thought process. You go, "Oh,
  343. 13:01that's true. It's true that my good
  344. 13:04example is super tight." And the
  345. 13:06opposite of super tight is verbose. So
  346. 13:09again, using these tools together, few
  347. 13:11shot prompting and chain of thought
  348. 13:13reasoning enables you to not only be
  349. 13:15able to create an example to emulate,
  350. 13:17but also a really good example to avoid.

  351. 3: Reverse Prompting

  352. 13:23The other technique that I think is kind
  353. 13:25of table stakes for collaborating well
  354. 13:27with AI is something called reverse
  355. 13:29prompting, which is basically asking the
  356. 13:32model to ask you for the information it
  357. 13:34needs. If you ask a model to write a
  358. 13:36sales email, it's going to make numbers
  359. 13:37up. And that can be frustrating to the
  360. 13:39uninitiated. You go, "Where did it get
  361. 13:41these sales numbers?" Well, here's my
  362. 13:42question. Did you give it your sales
  363. 13:44figures? How would it know? It's put
  364. 13:47placeholder text in and used its best
  365. 13:49guess. But if you reverse prompt the
  366. 13:52model and say at the end of your prompt,
  367. 13:55you know, help me write a sales email.
  368. 13:57Please walk me through your thought
  369. 13:58process step by step. Reference this
  370. 14:00good example and make it sound like
  371. 14:01that. and before you get started, ask me
  372. 14:04for any information you need to do a
  373. 14:06good job. The model will first walk you
  374. 14:08through its thought process and then
  375. 14:09instead of writing the email, it'll say,
  376. 14:11"I'm going to need the most recent sales
  377. 14:13figures to be able to write this email."
  378. 14:14Well, can you tell me how much you sold
  379. 14:16of this skew in Q2 last year? So, you
  380. 14:19basically give the model permission to
  381. 14:20ask you questions. This is part of the
  382. 14:22core actually of the teammate not
  383. 14:25technology paradigm. If you're working
  384. 14:27with a junior employee and you're
  385. 14:29sending them off on a task, what's one
  386. 14:30thing you're definitely going to say? If
  387. 14:32you have any questions, don't hesitate
  388. 14:33to ask me. Right? Any good manager,
  389. 14:36imagine a manager who says, "Don't ask
  390. 14:38me any questions." But sadly, AI in its
  391. 14:42desire to be a helpful assistant doesn't
  392. 14:44want to trouble us human with questions
  393. 14:46unless we give it permission to ask
  394. 14:48them.

  395. 4: Assigning a Role to AI

  396. 14:52Assigning a role is one of the most
  397. 14:54foundational techniques that you can
  398. 14:55leverage because it's effectively
  399. 14:58telling the AI where in its knowledge it
  400. 15:01should focus. So very simply, if you say
  401. 15:04you're a teacher, you're a philosopher,
  402. 15:06you're a reporter, you're a theatrical
  403. 15:09performer, molecular biologist, each of
  404. 15:12those titles triggers all sorts of deep
  405. 15:15associations with knowledge on the
  406. 15:18internet. you start to appreciate why
  407. 15:20simply giving a role helps because it
  408. 15:22starts to tell the AI where in your vast
  409. 15:26knowledge bank do I want you to draw
  410. 15:28information and make connections. So any
  411. 15:31one of them I would say is better than
  412. 15:33please review this correspondence. But
  413. 15:35better than just that prompt is saying
  414. 15:37I'd like you to be a professional
  415. 15:39communications expert. And if you have a
  416. 15:41favorite professional communications
  417. 15:42expert use them. I'd like you to take on
  418. 15:44the mindset of Dale Carnegie, the author
  419. 15:46of How to Win Friends and Influence
  420. 15:48Others. How would Dale Carnegie think
  421. 15:50about this? How do the principles that
  422. 15:52Dale Carnegie taught affect and
  423. 15:54influence and impact this
  424. 15:57correspondence? One of the simplest
  425. 15:58techniques that we teach at the Dh is
  426. 16:01trying on different constraints. One of
  427. 16:03the best ways you can solve a problem as
  428. 16:06a human is by forcing yourself to try on
  429. 16:08a bunch of different constraints. How
  430. 16:10would Jerry Seinfeld solve this problem?
  431. 16:12How would your favorite sushi restaurant
  432. 16:13solve this problem? How would Amazon
  433. 16:15solve it? How would Elon Musk? Anytime
  434. 16:18you make an association, you're
  435. 16:19colliding different information sources
  436. 16:22there. The same is true for an AI. An AI
  437. 16:25is basically making tons of connections
  438. 16:27through its own neural network. And by
  439. 16:29giving it a role, you're telling it
  440. 16:32where do you assume the best source of
  441. 16:35connection or collision is going to come
  442. 16:37from?

  443. 5: Roleplaying

  444. 16:41If I'm going to use AI to roleplay a
  445. 16:43difficult conversation, I typically
  446. 16:45think about kind of three different chat
  447. 16:48windows, so to speak, one is a
  448. 16:50personality profiler. Two is the
  449. 16:52character of the individual that I need
  450. 16:54to speak to, and then third is a
  451. 16:56feedback giver. I want to get objective
  452. 16:58feedback on the conversation. This I'll
  453. 17:00show you just how I would have a
  454. 17:02conversation with Chad GBT to prepare
  455. 17:05for a difficult conversation in my real

  456. If you’d like to try the Tough Convo Partner

  457. 17:07life. I'm just going to go into the
  458. 17:08tough conversation personality profiler
  459. 17:10and I'm going to say, "Hey, I'd love
  460. 17:11your help preparing for a conversation I
  461. 17:13need to have with my sales leader, Jim.
  462. 17:15He emailed me last night saying that he
  463. 17:17deserves commission on a deal that I
  464. 17:19know came through a different channel."
  465. 17:20And so, I'm just kind of giving a little
  466. 17:22bit of background. I will just upload
  467. 17:24that to the personality profiler. And
  468. 17:26what this one's been taught to do is I'm
  469. 17:28going to start with step one of the
  470. 17:29process, gather intelligence about the
  471. 17:31character and the scene. Right? I'm just
  472. 17:33going to look at the questions here and
  473. 17:34I'm going to I'm going to use my voice
  474. 17:36to answer them because it's a lot easier
  475. 17:37than using my fingers. Okay, first
  476. 17:39question. How would I describe Jim's
  477. 17:40communication style? Um, he's quite
  478. 17:42direct and confrontational. He's kind of
  479. 17:44typical East Coaster sarcastic. Well, I
  480. 17:47know that it came from our through our
  481. 17:49social team. There was a cold LinkedIn
  482. 17:51campaign that they ran and I know the
  483. 17:53CTO actually responded to that campaign.
  484. 17:55So, and then best case outcome of this
  485. 17:56conversation one I mean I'd like for Jim
  486. 17:58to kind of back down. I mean like
  487. 18:00near-term I want Jim to back down and
  488. 18:01agree that social team gets the
  489. 18:03commission. So it's kind of recapping
  490. 18:04its understanding and then it's giving
  491. 18:05me these instructions. It's saying to me
  492. 18:08the user copy paste the following
  493. 18:10instructions into a new Chad GPT window.
  494. 18:12So I've just copied those instructions.
  495. 18:14So now what I'm going to do is I'm
  496. 18:15opening Chad GPT and I'm just going to
  497. 18:17go and first thing I'm going to do is
  498. 18:19start a new conversation with Chad GBPT
  499. 18:21and I'm just going to paste those
  500. 18:22instructions in and I'm just going to
  501. 18:23paste them in. And what it should do is
  502. 18:25hopefully say got it. I'm Jim. So, I'm
  503. 18:28just going to go in and I'm going to
  504. 18:29start it. We're going to see how this
  505. 18:30goes. You guys ready?
  506. 18:34Hey, Jim. Um, do you have a second?
  507. 18:36Sure thing. I've always got a moment for
  508. 18:37a good chat. What's on your mind?
  509. 18:40Uh, yeah. I just wanted to clarify. It
  510. 18:42seems like um the commission that you
  511. 18:46emailed me about last night from the
  512. 18:47telco. I'm I'm afraid you might be
  513. 18:49mistaken there.
  514. 18:50Mistaken, huh? Well, I'm definitely open
  515. 18:52to hearing the details. If there's
  516. 18:55something I missed, let's lay it out on
  517. 18:56the table and figure it out.
  518. 18:58To be honest with you, I was a little
  519. 18:59bit caught off guard when you uh when
  520. 19:01you sent your email last night.
  521. 19:03All right, I hear you. I mean, if the
  522. 19:05social team really did the ground work
  523. 19:07to bring that lead in, then I definitely
  524. 19:10want to make sure credit goes where it
  525. 19:11belongs.
  526. 19:12So, if I hit X, I just I I end that
  527. 19:14conversation. What's interesting to me
  528. 19:16is I kind of feel like that wasn't
  529. 19:17totally gem. And either either that went
  530. 19:20well or you know I don't know what I did
  531. 19:22well. But one of the beautiful things
  532. 19:24about voice mode is now I've got the
  533. 19:25whole transcript here. I can just
  534. 19:27screenshot it. Right? So I can take
  535. 19:29screenshots of this conversation. I can
  536. 19:32put them into a new GPT window and I can
  537. 19:36get feedback and I can say, "Hey, tell
  538. 19:37me what I did well. Tell me what I did
  539. 19:38poorly." Or I can go back into my
  540. 19:40conversation history here and I could
  541. 19:43say, "Hey, I just had my first
  542. 19:45conversation with this version of Jim,
  543. 19:47but I feel like he kind of he was too
  544. 19:49agreeable to be honest with you, and so
  545. 19:51I'd like you to incorporate a little bit
  546. 19:53more edge into the instructions to the
  547. 19:57character that's going to be playing
  548. 19:58Jim, would you mind to update your
  549. 19:59instruction set?" Right? Because
  550. 20:01ultimately what you want is you want you
  551. 20:02want to have a realistic experience of
  552. 20:04the conversation. So, so now what I'll
  553. 20:06do is I'm actually going to open a new
  554. 20:07conversation and I'm going to paste the
  555. 20:09updated instruction set and let's see.
  556. 20:12Let's see what kind of gym I got here.
  557. 20:13Big voice, bigger results. I fight for
  558. 20:15what's mine. I don't appreciate being
  559. 20:17secondguessed. So, now I'm going to try
  560. 20:18again. Let's see. Hey, Jim. Do you have
  561. 20:20a minute to chat?
  562. 20:21Hey, absolutely. I've always got a
  563. 20:23minute. Oh, that one. Yeah, I I figured
  564. 20:25that might come up. I I definitely think
  565. 20:27I had a hand in that deal, so I'm
  566. 20:29curious to hear what you have to say
  567. 20:30about it.
  568. 20:31Yeah, really appreciate it. I mean, you
  569. 20:33what you and the sales team are doing is
  570. 20:34awesome. totally appreciate you. Um,
  571. 20:36that being said, I actually feel that
  572. 20:39attribution is really important that we
  573. 20:40get airtight here. And the truth is I I
  574. 20:42am fairly confident that deal came
  575. 20:44through the social team.
  576. 20:46All right, I get where you're coming
  577. 20:47from. I definitely appreciate you
  578. 20:49wanting to clear the air and to keep
  579. 20:51things fair. So, what kind of evidence
  580. 20:53do you have that it came through the
  581. 20:55social team?
  582. 20:55Well, so I've been through kind of two
  583. 20:57windows. one is profiling my
  584. 21:00conversation partner and then the second
  585. 21:02is having the conversation. All right.
  586. 21:04So, what I'm going to do here is I've
  587. 21:05opened my tough conversation greater GPT
  588. 21:08and I'm basically just going to upload
  589. 21:09photos of that last conversation.
  590. 21:13And what this GPT has been trained to do
  591. 21:15is basically give evaluate my
  592. 21:18conversation and then let me know how it
  593. 21:20went. Thanks for sharing the full
  594. 21:21transcript. My first step is to
  595. 21:23understand the objective. Step four,
  596. 21:24here's your grade. You got a 78 out of
  597. 21:26100. you succeeded in preserving trust
  598. 21:28and resolving the immediate issue. So, I
  599. 21:30can take all of these. I can even say,
  600. 21:32"Hey, would you give me a quick one
  601. 21:33pager of a handful of talking points
  602. 21:35that I should probably make sure not to
  603. 21:37forget in the order in which they're
  604. 21:39likely to emerge in this conversation
  605. 21:41based on the feedback you've given me."
  606. 21:42The AI will actually give me a really
  607. 21:45short kind of at a glance conversation
  608. 21:47guide that I can leverage if I want to
  609. 21:49try again. Right? Here's a one-pager.
  610. 21:51So, these are all great points. Now, I
  611. 21:53can bring them into the conversation. I
  612. 21:54actually I'd probably do this a couple
  613. 21:56times before having a real conversation
  614. 21:57with Jim. But the point is historically
  615. 22:01the only time I get feedback is after I
  616. 22:03have the real conversation with Jim.
  617. 22:05This is the first time in history and
  618. 22:07maybe I can get a friend to kind of go
  619. 22:09over talking points with me. But unless
  620. 22:11they're really close to gem or unless
  621. 22:13they're, you know, particularly
  622. 22:15imaginative and unless they're deeply
  623. 22:18knowledgeable of a bunch of feedback
  624. 22:19frameworks, they fall short of really
  625. 22:22preparing me in context for this
  626. 22:24specific situation in the specific
  627. 22:26conversation I need to have in a way
  628. 22:27that AI is able to help me. You can use
  629. 22:30this for any difficult conversation,
  630. 22:32whether it's a performance review, a
  631. 22:33salary negotiation, difficult feedback.
  632. 22:36It's a great way to basically get a
  633. 22:39flight simulator for a difficult
  634. 22:41conversation.
  635. 22:44The people who are the best users of AI
  636. 22:46are not coders. They're coaches. They
  637. 22:49aren't developers or software engineers.
  638. 22:52They're teachers and mentors and people
  639. 22:54who have learned to get exceptional
  640. 22:56output out of other intelligences. And
  641. 22:59so where could AI go? Well, it's really
  642. 23:02a function of who can get unleashed.
  643. 23:05Right now, the primary limitation is the
  644. 23:08limits of human imagination. And as we
  645. 23:11unleash and ignite and spark more humans
  646. 23:14imaginations, the kinds of applications
  647. 23:17that are possible or they're
  648. 23:18unthinkable, not because they're
  649. 23:20technologically impossible, but because
  650. 23:22they never occur to us personally. One
  651. 23:25of my favorite quotes is a Nobel
  652. 23:27Prize-winning economist named Thomas
  653. 23:29Shelling. He said no matter how heroic a
  654. 23:31man's imagination he could never think
  655. 23:33of that which would not occur to him. If
  656. 23:36you take as a premise that the
  657. 23:37imagination space as a function of what
  658. 23:40would occur to various individuals then
  659. 23:42as we equip different individuals what
  660. 23:44we can imagine collectively expands. In
  661. 23:47innovation studies has been called the
  662. 23:50adjacent possible for a long time. What
  663. 23:52is possible is just adjacent to what is.
  664. 23:55And as we increase adoption and increase
  665. 23:58fluency and competency and increasingly
  666. 24:01mastery of AI collaboration, then we're
  667. 24:04increasing the adjacent possible. And
  668. 24:07it's really important that you exercise
  669. 24:11through implementing some of the things
  670. 24:13you hear. And perhaps the most important
  671. 24:14thing you could do with this video is
  672. 24:16actually hit stop and do something
  673. 24:18that's already blown your mind.