Lesson 7: Effective prompting techniques (Deep Dive) | AI Fluency: Framework & Foundations Course — Transcript

Full transcript

1. 0:09[Music]
2. 0:12Let's explore one of the most practical
3. 0:14skills when working with AI. Crafting
4. 0:16effective prompts. This might sound
5. 0:18technical or complicated, and some
6. 0:20guides certainly make it seem that way,
7. 0:22but at its heart, it's surprisingly
8. 0:24straightforward. Prompting is simply how
9. 0:26we apply this course's description
10. 0:27competency in practice. clearly
11. 0:30communicating what we want, how we want
12. 0:31it done, and how we want to interact
13. 0:33with our AI assistant throughout the
14. 0:35entire process. Think of prompting like
15. 0:37explaining a task to a helpful new
16. 0:39colleague who's eager to assist, but
17. 0:41needs clear directions and expectation
18. 0:43setting to do their best work. We'll be
19. 0:45using Claude throughout this section,
20. 0:47but these tips can be carried over to
21. 0:49many other AI systems. You might have
22. 0:52heard the term prompt engineering tossed
23. 0:54around. Prompt engineering is simply the
24. 0:56practice of designing effective
25. 0:58instructions for AI systems like Claude.
26. 1:01It's about crafting your questions and
27. 1:02providing context in ways that help AI
28. 1:04assistants understand exactly what you
29. 1:07want. What's fascinating is that
30. 1:09effective prompting blends familiar
31. 1:11human communication skills with a few
32. 1:13considerations specific to AI. Many
33. 1:16principles that make for good human
34. 1:17conversation, such as being clear,
35. 1:19providing relevant context, and giving
36. 1:21concrete examples, also apply when
37. 1:23working with AI. Yet, there are
38. 1:25differences, such as being more explicit
39. 1:27about things humans could naturally
40. 1:28infer, and accommodating the AI's
41. 1:31limited context window, and sometimes,
42. 1:33depending on the AI you're working with,
43. 1:35using specific formatting that machines
44. 1:37can easily process. As AI assistants
45. 1:40continue to evolve, prompting best
46. 1:42practices evolve, too. What works with
47. 1:44today's AI systems may be different from
48. 1:46what works with tomorrow's.
49. 1:48Experimentation is key to discovering
50. 1:50what works best for your specific needs.
51. 1:52In this video, we'll mainly explore six
52. 1:54foundational prompting tips that will go
53. 1:56a long way toward helping you
54. 1:57effectively communicate and collaborate
55. 1:59with Claude and other AI systems. They
56. 2:01are give Claude context, show examples
57. 2:04of what good looks like, specify output
58. 2:07constraints, break complex tasks into
59. 2:10steps, ask Claude to think first, and
60. 2:13define Claude's role, style, or tone.
61. 2:15The first principle is simple but
62. 2:17powerful. Be specific and clear about
63. 2:19what you want, why you want it, and
64. 2:21perhaps most surprisingly, who you are.
65. 2:24Let's take a simple prompt. Tell me
66. 2:26about climate change. How can we improve
67. 2:29this by giving Cloud more context? A
68. 2:31more specific contextrich version could
69. 2:33look like, explain three major impacts
70. 2:36of climate change on agriculture in
71. 2:38tropical regions with examples from the
72. 2:40past decade. Our baseline prompt was
73. 2:43vague and leaves clawed guessing about
74. 2:44our interests, level of knowledge, and
75. 2:46the depth of detail we're looking for,
76. 2:48such as geography and time span. We can
77. 2:50even add more context by providing
78. 2:52information not just about what we're
79. 2:54looking for, but why we're asking and
80. 2:56how we'll be using that information that
81. 2:58Claude gives us. Now, our prompt looks
82. 3:00like this. Explain three major impacts
83. 3:02of climate change on agriculture in
84. 3:04tropical regions with examples from the
85. 3:06past decade. I'm preparing for a job
86. 3:09interview at an agricultural research
87. 3:10lab in Indonesia. I have a degree in
88. 3:13ecology, but no specific knowledge on
89. 3:15climate change. write a summary of key
90. 3:17concepts that would help me speak
91. 3:18intelligently in the interview. All this
92. 3:21added context helps tailor Claude's
93. 3:23response to your specific situation and
94. 3:25knowledge level. This kind of background
95. 3:26information is something we naturally
96. 3:28provide in human conversations, but
97. 3:30might forget to include when talking
98. 3:31with Claude. Sometimes showing is better
99. 3:34than telling. Providing examples of the
100. 3:36kind of output you're looking for can be
101. 3:38incredibly effective. This is sometimes
102. 3:40called fshot prompting or nshot
103. 3:42prompting in technical circles where n
104. 3:44is the number of examples given but it's
105. 3:47really just about showing the AI
106. 3:48examples for it to emulate. For
107. 3:50instance, take the following prompt.
108. 3:52Please convert this technical statement
109. 3:54to plain language. The platform
110. 3:56implements end-to-end encryption
111. 3:57protocols to safeguard data integrity.
112. 4:00Clog may already be able to do this to
113. 4:02your satisfaction. So we definitely
114. 4:04recommend you just try first without
115. 4:06examples and see where it leads you. But
116. 4:08let's say you have a very specific style
117. 4:10you want Claude to follow, and it's
118. 4:12harder to explain than to give examples.
119. 4:15Your refreshed prompt could look
120. 4:16something like this. Here are two
121. 4:19examples of how to convert technical
122. 4:20jargon into plain language. Original,
123. 4:23the quantum algorithm exhibits quadratic
124. 4:25speed up. Plain, the new method solves
125. 4:27problems roughly twice as fast as
126. 4:29previous methods. Original, the
127. 4:32interface leverages intuitive design
128. 4:33paradigms. Plain, the design is easy to
129. 4:36understand and use. Now, please convert
130. 4:39this complex technical manual to plain
131. 4:41language. When providing examples, aim
132. 4:43to cover the full diversity of possible
133. 4:45prompts, such as examples that cover
134. 4:47different cases or styles. This helps
135. 4:50Claude better understand the broad range
136. 4:51of the pattern you want it to follow.
137. 4:53Being clear about output constraints,
138. 4:55such as the desired format and length of
139. 4:57Claude's response, or the language you
140. 4:59want Claude to code in, or the color of
141. 5:01the buttons on the web page you want
142. 5:03Claude to design, also helps ensure you
143. 5:05get exactly what you need. Here's an
144. 5:08example of clear and detailed
145. 5:09description to ensure Claude delivers
146. 5:11exactly what you're looking for. Create
147. 5:14a clean, modern, single page art
148. 5:16portfolio website. Include these main
149. 5:18sections: hero, about me, skills,
150. 5:19portfolio, projects, experience, and
151. 5:21contact. Make the navigation menu sticky
152. 5:24and responsive with hamburger menu on
153. 5:25mobile. Use a sunset color palette and
154. 5:28add a dark light mode toggle in the
155. 5:30navigation. Guidance like this helps
156. 5:32cloud structure its response to match
157. 5:34your expectations. When you have a
158. 5:36complicated request, breaking it down
159. 5:38into smaller steps helps Cloud follow
160. 5:40your thinking and deliver better
161. 5:42results. Think about it this way. If you
162. 5:45ask a friend to do something for you
163. 5:46without specifying how, there's a chance
164. 5:48that they may not do it the way you
165. 5:50intended them to. We've all been there.
166. 5:52Listing out task steps ensures that
167. 5:54Claude follows the process you want to
168. 5:56in order to accomplish its task. This is
169. 5:59sometimes called chain of thought
170. 6:00prompting. For example, instead of
171. 6:02asking Claude to analyze this quarterly
172. 6:04sales data, you might say, "I'd like to
173. 6:06analyze this quarterly sales data.
174. 6:08Please approach this by looking through
175. 6:10our sales records to identify the top
176. 6:11performing products, comparing current
177. 6:14quarter results to the previous quarter,
178. 6:16highlighting any unusual trends or
179. 6:17patterns, and then suggesting possible
180. 6:19reasons for these trends. By default,
181. 6:21you may not need to do this, especially
182. 6:23for tasks that are relatively
183. 6:24straightforward. Furthermore, modern
184. 6:27reasoning models or extended thinking
185. 6:29models are increasingly capable of
186. 6:31performing step-by-step reasoning on
187. 6:33their own, but you can still guide this
188. 6:36process to ensure it aligns with your
189. 6:37needs. The more variance there is in
190. 6:40ways to execute the task well, or the
191. 6:42more that proper task execution relies
192. 6:44on experience and knowledge you've
193. 6:45gained as a domain expert, the more you
194. 6:48should consider taking the time to
195. 6:49translate that knowledge into Claude.
196. 6:51Relatedly, sometimes it can be helpful
197. 6:53to explicitly give AI assistants like
198. 6:55Claude space to work through its process
199. 6:57first before executing its task. This
200. 7:00approach helps Claude produce more
201. 7:02thorough and well-considered responses.
202. 7:04For example, you can add this to your
203. 7:06prompt. Before answering, please think
204. 7:08through this problem carefully. Consider
205. 7:11the different factors involved,
206. 7:12potential constraints, and various
207. 7:14approaches before recommending the best
208. 7:16solution. As I mentioned, modern
209. 7:18reasoning or extended thinking models by
210. 7:20default think before acting. But if
211. 7:22you're working with an AI assistant that
212. 7:24does not think first by default, you can
213. 7:26still prompt the AI to do so. I want to
214. 7:28note the importance of giving the AI
215. 7:30assistant space to think before doing
216. 7:32its task, not after. If you want that
217. 7:35thinking to increase the quality of the
218. 7:37AI's work, just like how having space to
219. 7:39think before you act is different than
220. 7:41acting first, then being asked to
221. 7:43explain your thinking afterwards. As a
222. 7:45side benefit, this also allows you to
223. 7:47better see where the AI assistant might
224. 7:49be going astray and thus where you could
225. 7:51hone your description competency further
226. 7:53by providing more guidance. Specifying
227. 7:56how you want cloud to communicate and
228. 7:58behave can significantly change how it
229. 8:00approaches a task. By specifying the
230. 8:02level of expected expertise, the
231. 8:04perspective you want it to take, or its
232. 8:06communication style, you can guide both
233. 8:08Claude's interaction with you and the
234. 8:10final result of what it produces. Simply
235. 8:12put, who do you want the AI to act as?
236. 8:15For example, take this prompt. Please
237. 8:17explain how rainbows form from the
238. 8:19perspective of an experienced science
239. 8:21teacher speaking to a bright 10-year-old
240. 8:23who's interested in science. This is
241. 8:25also a good way to brainstorm or get
242. 8:27feedback. You can specify a general role
243. 8:29or even ask Claw to take on the persona
244. 8:31of a specific figure, such as Richard
245. 8:32Fineman, when asking for physics
246. 8:34explanations. Here's another example. As
247. 8:37a UX design expert, review this website
248. 8:40wireframe and suggest three improvements
249. 8:42focusing on user navigation and
250. 8:43accessibility. Perhaps the most powerful
251. 8:45technique is asking Claude to help
252. 8:47improve your prompt. When you're not
253. 8:49sure how to ask for something or how to
254. 8:51improve your prompt, describe to Claude
255. 8:53your issue or situation and ask it to
256. 8:55make your prompt better or write your
257. 8:56prompt for you. I'm trying to get you,
258. 8:59Claude, to help me with goal. I'm not
259. 9:01sure how to phrase my request to get the
260. 9:04best results. Can you help me craft an
261. 9:06effective prompt for this? Here's where
262. 9:08Claude and other AI assistants may vary
263. 9:10most in terms of performance. So, we
264. 9:12suggest you experiment with different
265. 9:13models as part of practicing delegation.
266. 9:17Effective prompting is iterative and
267. 9:19experimental. AI systems and best
268. 9:21practices are constantly evolving. So,
269. 9:23what works today may change tomorrow.
270. 9:25Your first attempt won't always yield
271. 9:27the perfect result, and that's expected.
272. 9:29When a response isn't quite what you
273. 9:31need, try refining your approach by
274. 9:33playing around with any of the
275. 9:34techniques we mentioned, such as add
276. 9:37more specificity or context. Provide
277. 9:40examples of your desired output. Break
278. 9:42the task into smaller steps and try a
279. 9:45different technique or combination of
280. 9:47techniques. You can also ask for
281. 9:49variations such as, "Can you give me
282. 9:50three different versions of this?" You
283. 9:53can request different formats such as,
284. 9:54"Instead of a paragraph, could you
285. 9:56present this in an interactive
286. 9:58artifact?" Note that artifacts are a
287. 10:00unique way that Claude can create
288. 10:01outputs that may be easier to understand
289. 10:03or more interesting to digest. You can
290. 10:05also check confidence, such as for
291. 10:07factual questions, you can ask, "How
292. 10:09confident are you about this answer?"
293. 10:12You can also reset the conversation
294. 10:13entirely. Sometimes starting a fresh
295. 10:15conversation gives better results than
296. 10:17trying to correct the conversation
297. 10:19that's gone off track. Use each
298. 10:21interaction as feedback to improve your
299. 10:23next prompt. Over time, you'll develop
300. 10:25an intuition for how to communicate
301. 10:27effectively with all AI systems. As you
302. 10:30apply these techniques in practice,
303. 10:31here's some guidance to recap. Some
304. 10:34patterns consistently work well.
305. 10:36Starting with a clear task overview
306. 10:37statement, including format
307. 10:39specifications and examples, setting
308. 10:41explicit constraints or requirements,
309. 10:43providing rich and relevant background
310. 10:45information, and common mistakes to
311. 10:47avoid are assuming that Claude can read
312. 10:49your mind, or overloading a single
313. 10:52prompt or conversation with multiple
314. 10:53unrelated tasks, being too vague about
315. 10:56what success looks like, and not
316. 10:57providing feedback on previous
317. 10:59responses. To recap, effective
318. 11:01communication with AI systems like
319. 11:03Claude combines timeless human
320. 11:04communication principles with AI
321. 11:06specific techniques. The approaches
322. 11:08we've covered will serve you well across
323. 11:10different AI systems. These six
324. 11:12principles together with the secret
325. 11:13weapon of asking Cloud for help form a
326. 11:16solid toolkit for applying the
327. 11:17description competence to your AI
328. 11:19interactions. Iteration and practice
329. 11:21here is the key to Swift improvement and
330. 11:24mastery. Remember that prompt
331. 11:26engineering is an evolving practice. As
332. 11:28models improve, some specific techniques
333. 11:30become less necessary. However, these
334. 11:33principles of good communication are
335. 11:35still relevant even if the way we apply
336. 11:37them changes. Maintain a spirit of
337. 11:39experimentation and adapt your approach
338. 11:42based on your results.