Top 6 AI Productivity Tips for Work! — Transcript

Full transcript

1. 0:00If you're tired of learning AI tips that
2. 0:01go outdated almost immediately, then
3. 0:04you're in the right place. I've been
4. 0:05using every major AI tool since Chachi
5. 0:07PT first came out, and the techniques
6. 0:09that actually last are the ones I've
7. 0:11taught at institutions like Google, PWC,
8. 0:13and Stanford. In this video, I'm
9. 0:15covering the top six practical AI tips
10. 0:17that will still work years from now,
11. 0:19grouped under three categories. Waste
12. 0:22fewer tokens, do less manual work, and
13. 0:24get AI to learn your preferences over
14. 0:27time. Let's get started. Kicking things
15. 0:29off with tip number one, the markdown
16. 0:31shortcut. Basically, the AI wastes a lot
17. 0:34of time and tokens whenever you upload a
18. 0:36PDF because it's processing the page
19. 0:38layout along with the text. Whereas, a
20. 0:41markdown file lets the AI get straight
21. 0:43to the information by stripping out
22. 0:45unnecessary visuals and formatting. For
23. 0:47example, let's say I'm working with this
24. 0:49incredibly interesting and frankly
25. 0:51speaking addictive 42page text guide.
26. 0:54Instead of uploading this entire PDF
27. 0:55onto Tragic PT, for example, I would
28. 0:58instead press command and control A to
29. 1:00select all the text. Command control C
30. 1:02to copy, go back to Chrome or your
31. 1:05browser of choice. Type in doc new to
32. 1:08open up a blank Google [music] doc.
33. 1:10Paste the text. As you can see, there's
34. 1:1248 pages of this. And then go to file,
35. 1:16download markdown. And now after this
36. 1:20gets downloaded, we can see that instead
37. 1:22of a 661 kilob PDF file, we now have a
38. 1:2894 kilobyte lightweight markdown
39. 1:32document that AI can process much faster
40. 1:35while using way fewer tokens. But what
41. 1:38if you don't want to repeat that manual
42. 1:39process every single time? In that case,
43. 1:41[music] I can create an empty folder on
44. 1:43my desktop, name it something like
45. 1:45personal finances and drop the PDF
46. 1:47inside. Then after I mount that same
47. 1:50folder in the desktop version of Kajipt
48. 1:53Cloud or Gemini, which you should be
49. 1:55definitely using over the web versions,
50. 1:56by the way, I simply instruct the AI to
51. 1:59extract parts of the PDF that are
52. 2:01relevant to me into a markdown file. And
53. 2:04we'll just fast forward this part. Here
54. 2:06we can see the AI created a markdown
55. 2:08file that's over 90% smaller than the
56. 2:11original. So now with the same folder
57. 2:13connected, the AI can use that smaller
58. 2:15reference file for all future tax
59. 2:18related tasks, saving time and tokens
60. 2:21with each use. And when you work with
61. 2:23hundreds of dense files like I do, those
62. 2:25savings add up real quick. As a rule of
63. 2:28thumb, this applies to any file that's
64. 2:30mostly text. Meaning you always want to
65. 2:32give AI the plain text version. Whether
66. 2:34that's a markdown file instead of a PDF
67. 2:36or Word doc or a CSV instead of a messy
68. 2:40Excel spreadsheet full of merge cells
69. 2:42and extra tabs. Pro tip, if you're still
70. 2:44using projects in the web version of
71. 2:46Chachi T or Claude, you want to upload
72. 2:48both the original PDF and the
73. 2:50lightweight markdown file as context
74. 2:52files. And within the project
75. 2:55instructions, let me just quickly find
76. 2:56it here. You want to tell the AI to read
77. 2:59the markdown files by default and only
78. 3:02open the PDF files when it needs more
79. 3:05information. And if you're using one of
80. 3:06the desktop apps with a home folder
81. 3:08mounted like I am here, I recommend
82. 3:11storing all the original PDFs in a
83. 3:14source PDFs folder and then adding a
84. 3:17rule in your instructions file, which is
85. 3:19agents.md for cach users or claw.md for
86. 3:23clawed users. You want to add a rule
87. 3:26instructing the AI to only read the
88. 3:28original PDF as a fallback [music]
89. 3:31option. By the way, if setting up Chatt
90. 3:33or Claude's desktop app with folders and
91. 3:35instructions like this is completely new
92. 3:37to you, don't worry because my free
93. 3:40Claude toolkit walks you through what to
94. 3:42do step by step. And as one subscriber
95. 3:46put it, they didn't expect to get so
96. 3:48much out of a free email list. So, link
97. 3:50down below. Next up, the second tip for
98. 3:52spending less on tokens, the folder map.
99. 3:54Diving right into an example, let's say
100. 3:56I have over a dozen or hundreds of files
101. 3:58in this folder, including interview
102. 4:00notes, survey results, and even some raw
103. 4:03screenshots. If my task only needs a few
104. 4:06of these files, the AI would waste a lot
105. 4:09of time and tokens opening and reading
106. 4:11all of them. Instead, I'd mount the
107. 4:13folder and ask the AI to read every file
108. 4:15in this folder just once to create an
109. 4:18index listing each file's location, a
110. 4:22one-s sentence description, and when to
111. 4:24use it. And we're just going to fast
112. 4:26forward here. Now that we have this
113. 4:27index.md file, any AI with access to
114. 4:30this will know what each file is about.
115. 4:33So the next time I ask about the most
116. 4:36common customer complaints, for example,
117. 4:38the AI reads the index first to identify
118. 4:41the interview notes containing negative
119. 4:43sentiment and only reads those files,
120. 4:47saving the time and tokens it would have
121. 4:48spent opening everything else. Pro tip,
122. 4:50you can create a schedule task that
123. 4:52instructs the AI to scan your folder
124. 4:55periodically so that when you add or
125. 4:57remove files, the AI updates the index
126. 5:00automatically to reflect the latest
127. 5:02changes. This works for any large
128. 5:04collection of files, by the way, like
129. 5:05thousands of receipts for expense claims
130. 5:07or hundreds of raw photos for my YouTube
131. 5:10thumbnails. Moving on to the first tip
132. 5:12for doing less manual work, prompt
133. 5:14shortcuts. Diving right into a simple
134. 5:16example, I'm writing an important email
135. 5:18to senior leadership, but the wording in
136. 5:20my first draft feels awkward. And so, in
137. 5:23a new chat, I type GPT [music]
138. 5:26rephrase. and my text expander instantly
139. 5:30paste this saved prompt asking for three
140. 5:32clearer and more conversational
141. 5:34variations. So I just paste my draft
142. 5:37underneath, hit send, and skipping ahead
143. 5:39here, we can see the three improved
144. 5:42versions. Now, many of you have asked me
145. 5:45whether text expanders are still
146. 5:46relevant since skills work in a similar
147. 5:48way. And the rule of thumb is to use a
148. 5:51text expander for prompts you need to
149. 5:52tweak before sending and use skills for
150. 5:56recurring multi-step tasks that should
151. 5:58follow the exact same steps every time.
152. 6:01For example, this is my inbox triage
153. 6:03skill. And technically, I could save
154. 6:05this entire skill, all 25,533
155. 6:09words of it in my text expander, but
156. 6:11that would be silly since number one, I
157. 6:13can just run it by typing for/inbox
158. 6:16triage like so. And second, if I need to
159. 6:19update the skill when a new model drops,
160. 6:21for example, I can simply ask the AI to
161. 6:24edit the skill directly by saying update
162. 6:26the skill for the latest Opus model, for
163. 6:28example, instead of trying to fix it in
164. 6:30the text expander app myself. So when is
165. 6:33it better to use text expanders? [music]
166. 6:35Here I can type colon GBT AI prompt to
167. 6:40paste a prompt template designed to
168. 6:43write a prompt optimized for a specific
169. 6:45model. And as you can see, the model
170. 6:47name is a placeholder. So I can just
171. 6:49replace this with like Astra or Opus or
172. 6:51Argan, [music]
173. 6:52right? And since the whole prompt is
174. 6:55right in front of me, I can tweak
175. 6:57anything else before I send it, which is
176. 6:59obviously a lot harder to do with a
177. 7:02skill. I personally use the paid version
178. 7:04of Alfred as my text expander app, but
179. 7:06free alternatives include Raycast for
180. 7:08Mac users and Beefex for Windows users.
181. 7:11As a quick recap, you should be using
182. 7:13both [music] text expanders and skills.
183. 7:16And over time, you'll develop a sense of
184. 7:18which tasks belong in a skill and which
185. 7:20prompts belong in your text expander.
186. 7:22All right, so just now I mentioned that
187. 7:24it's best to update your skills after a
188. 7:26new model comes out. And the reason for
189. 7:28that is because the instructions you
190. 7:30save for older models can sometimes hold
191. 7:33back the newer and more capable ones.
192. 7:35And I actually learned this tip from
193. 7:37Ruben Hassid, who I actually follow on
194. 7:38LinkedIn to learn practical AI tips. So,
195. 7:41it feels like a bit of a full circle
196. 7:43moment to be sharing a free guide he put
197. 7:45together with today's sponsor, HubSpot,
198. 7:47called Use Claude like an AI expert. And
199. 7:49even though the title says Claude, the
200. 7:51templates and tips inside work with
201. 7:53ChateBT and Gemini as well. And I highly
202. 7:56recommend going over the expert insight
203. 7:58section in particular since it really
204. 8:01helps you rethink how you work with AI.
205. 8:04For example, Reuben points out a lot of
206. 8:06people don't think they're using AI
207. 8:07agents when in reality, even a simple
208. 8:10scheduled task already counts as one.
209. 8:12And a few of his other insights, like
210. 8:14how not to overthink which model to use
211. 8:16with whatever level, tie in nicely with
212. 8:18the bonus lightning round I'm saving for
213. 8:20the end of this video. But if you want
214. 8:22the full set, it's all in the guide.
215. 8:24Click the first link in the description
216. 8:25to get the entire guide for free. And
217. 8:27thank you HubSpot for sponsoring this
218. 8:28video. Next up, we have my personal
219. 8:30favorite, the intent check. If we're
220. 8:32being honest, most of us oneshot our
221. 8:35tasks. Meaning, we type out a request
222. 8:37once and hope the AI nails it on the
223. 8:39first try. And because models have
224. 8:41gotten so powerful, that usually works
225. 8:43fine for low stakes tasks where small
226. 8:46mistakes don't really matter. But for
227. 8:48high stakes requests, research from
228. 8:50Google, Anthropic, and OpenAI all agree
229. 8:55that getting the AI to clarify the goal
230. 8:57and plan the work before it starts can
231. 8:59drastically improve the output. And the
232. 9:01intent check is the simplest way to do
233. 9:03this. At the start of an important task,
234. 9:05have the AI repeat your underlying
235. 9:07intent back to you and propose a plan
236. 9:09for your review. So you can course
237. 9:11correct before it does any of the work.
238. 9:14In this first example, I have a brief
239. 9:15and three proposals saved in a planning
240. 9:18folder. And with the folder attached, I
241. 9:20asked the AI, "Three wedding planning
242. 9:22agencies sent over their proposals.
243. 9:24Which one should I hire?" And if we fast
244. 9:27forward a bit here, we get exactly what
245. 9:29we asked for. a first choice
246. 9:31recommendation. But notice how the AI
247. 9:34only explains its rationale in words
248. 9:36instead of scoring each agency against a
249. 9:38clear rubric, which is something I would
250. 9:40have wanted to see for a decision this
251. 9:42big. So, in a new chat, I have the same
252. 9:45folder mounted, and the prompt starts
253. 9:46off the exact same way, but [music] this
254. 9:48time I type GPT intent to paste my
255. 9:52intent check prompt, [music] which I'll
256. 9:54link down below. that instructs the AI
257. 9:56to repeat its understanding of my ask
258. 9:59and propose a plan for me to review
259. 10:01before it does any work. And let's skip
260. 10:05ahead. You'll notice that this time,
261. 10:07instead of jumping straight to
262. 10:08recommendation, the AI first repeats
263. 10:10back what I'm actually trying to do,
264. 10:12including my non-negotiables, like a
265. 10:143:00 a.m. afterparty, then lays out its
266. 10:16three-step plan, so I can see exactly
267. 10:19how it'll approach my request before it
268. 10:21does any work. And because it listed
269. 10:23every step, right, I can see the plan
270. 10:25doesn't include a scorecard. So I just
271. 10:27reply, hey, the plan is approved
272. 10:29provided that you also create a
273. 10:31scorecard with five criteria. So I can
274. 10:33see how the agencies stack up against
275. 10:35each other. And if we send this off and
276. 10:37fast forward, this weighted scorecard
277. 10:40shows exactly how the AI reached its
278. 10:42recommendation. [music]
279. 10:43And I can now check whether the criteria
280. 10:46and weights match my own preferences.
281. 10:48So, as a rule of thumb, use this prompt
282. 10:50that I'll link down below before any
283. 10:51high impact task. Since seeing the AI's
284. 10:54plan up front makes it easy to spot the
285. 10:56context gaps only we humans can [music]
286. 10:59fill and course correcting early is a
287. 11:02lot cheaper than realizing halfway
288. 11:03through the AI completely misunderstood
289. 11:05your intent. Rounding out the tips for
290. 11:07doing less manual work, the spare key.
291. 11:10Basically, most apps we use have two
292. 11:12entrances. the human-friendly interface
293. 11:14where we log in and click around and a
294. 11:17separate entrance built for software
295. 11:19called [music] an API which used to be
296. 11:21mainly for developers since it involves
297. 11:24coding. [music] Luckily for the rest of
298. 11:25us, AI can now write that code so we can
299. 11:27simply give the AI an API key for that
300. 11:30second entrance and have it do all the
301. 11:32things that used to require a developer.
302. 11:35For example, here's my website and it
303. 11:36runs on an open-source platform called
304. 11:38Ghost. And since I'm not technical at
305. 11:41all, I was stuck with whatever front-end
306. 11:44settings [music] Ghost gives me and its
307. 11:46users, right? Meaning, if I wanted to
308. 11:48make a larger revamp of my website,
309. 11:52added to pay technical freelancers to do
310. 11:54them for me now, I can just create an
311. 11:56API key, save it in a secure location
312. 11:59like on my Max keychain, and then
313. 12:01connect it to my AI tool of choice and
314. 12:04just describe the change I want in plain
315. 12:06English. And using the wayback machine,
316. 12:09you can actually see this is how my
317. 12:11newsletter page used to look, right?
318. 12:13Pretty boring. And I was stuck with like
319. 12:15the default and kind of ugly carousel
320. 12:17card here. And this is how it looks now.
321. 12:21I just described the vibe I [music]
322. 12:23wanted. And the AI even added a
323. 12:25scrolling testimonial wall of love.
324. 12:28[snorts] And I even had the AI audit my
325. 12:30entire website to fix typos and
326. 12:33formatting issues across hundreds of
327. 12:35pages. This is especially useful when a
328. 12:37tool does not have a native plug-in or
329. 12:40MCP connector you can use. And even when
330. 12:42it does, an API key usually gives you
331. 12:45broader access, so you can do a lot more
332. 12:48with it. Pro tip: API keys are a lot
333. 12:50more common than you think. So, you can
334. 12:52just search your tool's name like Kit
335. 12:53followed by the word API, and you'll
336. 12:56usually find a page telling you exactly
337. 12:58how to set it up on that platform. Just
338. 13:01remember the most important rule. Never
339. 13:04ever paste your API key directly in your
340. 13:07prompt. Instead, [music] have your AI
341. 13:09walk you through how to store the API
342. 13:11key securely in your Mac keychain or
343. 13:13Windows credential manager where you
344. 13:15type it in yourself so your AI can use
345. 13:18the key without it ever showing up in
346. 13:20the chat. Moving on to the first and
347. 13:22only tip in our third category, tool
348. 13:24rule books. This tip gets its own
349. 13:26category because AI agents can be pretty
350. 13:29intimidating for non-technical users.
351. 13:31And this tip is the easiest way to start
352. 13:34learning about AI agents like traypic
353. 13:36work, cloud code, and Google
354. 13:37anti-gravity since it teaches you
355. 13:39exactly how agents learn your
356. 13:41preferences over time. For example, here
357. 13:44I'm in the web version of Claude chat
358. 13:45with Gmail connected and I ask it to
359. 13:47draft a reply to Fireship's email
360. 13:50saying, "I'd love to help with this
361. 13:51horse Tinder project." [music] And fast
362. 13:53forwarding a bit, we can see Claude
363. 13:55wrote a decent reply, but it's longer
364. 13:57than what I would have written and it
365. 13:58signs off with my name. So over in
366. 14:01Gmail, I still have to trim it [music]
367. 14:03and delete the signoff since Gmail
368. 14:06already adds my signature by default.
369. 14:08Now, here's that exact same prompt in
370. 14:10the Claw desktop app and with a folder
371. 14:12mounted that contains my Gmail rule
372. 14:14book. And skipping ahead a bit, we can
373. 14:16see the output is completely different.
374. 14:19It's short and to the point, the first
375. 14:21sentence is bolded [music] and there's
376. 14:23no sign off, so I can paste it straight
377. 14:25into Gmail and hit send. The broader
378. 14:27principle is simple. Connecting to tools
379. 14:30like Gmail, Slack, Notion, or Google
380. 14:32Calendar only gives the AI access
381. 14:34[music]
382. 14:35to those tools. Meaning, you still need
383. 14:37a rulebook for the AI to use those tools
384. 14:39the way you would. And setting one up is
385. 14:42a lot easier than you think. I first
386. 14:44create an empty folder on my desktop,
387. 14:46and I mount that empty folder in Claude,
388. 14:49for example, and then I can just ask
389. 14:51Claude in plain English to create two
390. 14:53files. a claw.md file which is the main
391. 14:56instruction file Claude reads at the
392. 14:58start of every session in chatbt that's
393. 15:00called agents.mmd and a Google calendar
394. 15:02rulebook with one rule events always
395. 15:06default to 45 minutes and we're just
396. 15:08going to fast forward this [music] part.
397. 15:10All right, the two files have been
398. 15:11created and opening up the cloud. D
399. 15:13file, we see instructions that tell
400. 15:16Claude to read the calendar rule book
401. 15:18whenever calendar comes up. And the
402. 15:21calendar rulebook contains the rule I
403. 15:24asked for. And putting this straight to
404. 15:26the test, I start a brand new chat with
405. 15:28the same folder mounted and ask Claude
406. 15:31to create an event at 10:00 a.m.
407. 15:33tomorrow. And skipping ahead just a bit,
408. 15:35we can see it creates a 45minute event
409. 15:38instead of the usual hour. Obviously,
410. 15:40that was a very simple example for
411. 15:42illustration purposes. So, a more
412. 15:44practical calendar rule to add would be
413. 15:46telling the AI to keep private [music]
414. 15:48notes out of the event description
415. 15:50whenever other people are invited. And
416. 15:53small improvements like that make a
417. 15:55massive difference over time. As a rule
418. 15:57of thumb, you want to create one rule
419. 15:59book for each tool you have connected.
420. 16:02And pro tip, you don't have to come up
421. 16:04with every rule yourself. Just ask your
422. 16:06AI to review your connected tools, draft
423. 16:09a rulebook for each one based on how you
424. 16:11already use them, and ask you about
425. 16:13anything that's missing. As you can see,
426. 16:15my actual Jeffos workspace contains an
427. 16:18entire folder of these conventions
428. 16:21files, right? [music] And these are all
429. 16:23separate, meaning the AI would only load
430. 16:25the ones a task actually needs, meaning
431. 16:28I get faster and just better outputs. By
432. 16:30the way, if you want this exact setup
433. 16:32for yourself, I share my entire system
434. 16:34in the Cowork Academy, so you can skip
435. 16:36months of trial and error. It arrives
436. 16:3890% built, so you can have it up and
437. 16:40running in under 5 hours. [music] Link
438. 16:42down below. Now, for the bonus lightning
439. 16:44round I promised, four quick tips, and
440. 16:46this is going to go fast. First,
441. 16:48organize your feedback to AI using
442. 16:50hashtags. For example, if the AI drafted
443. 16:52a presentation for you, you can put # #
444. 16:56slide [music] 2 above your comments on
445. 16:59slide 2 and triple #chart
446. 17:03underneath it for comments on just that
447. 17:06slides chart. That way, the AI knows
448. 17:09exactly which comments apply to which
449. 17:11parts of its output. Second, one good
450. 17:14example beats an overengineered prompt.
451. 17:16For instance, paste last month's report
452. 17:18and say, "Write this month's update in
453. 17:20the same format, and the AI will match
454. 17:23the structure, length, and tone much
455. 17:25more closely than even the most detailed
456. 17:27[music] prompt." Third, approve a small
457. 17:29sample before the whole batch. If you're
458. 17:31asking AI to rename 200 files, have it
459. 17:34do the first five, check them, then let
460. 17:37it run the rest. So, if something's off,
461. 17:39you catch it once instead of fixing it
462. 17:41200 times. And finally, let one AI check
463. 17:44another's work before you [music] do.
464. 17:46When chatting BT, write something
465. 17:48important like a report for your boss.
466. 17:50Paste it into Claude or Gemini or Kimmy
467. 17:52and ask it to flag mistakes and gaps so
468. 17:55you catch problems before you spend time
469. 17:57reviewing the output yourself. If you
470. 17:59enjoyed these tips, you might want to
471. 18:00check out this tutorial on how I use AI
472. 18:03to manage my files. See you on the next
473. 18:05video. In the meantime, have a great
474. 18:07one.