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

[00:09] [Music]

[00:12] Let's explore one of the most practical

[00:14] skills when working with AI. Crafting

[00:16] effective prompts. This might sound

[00:18] technical or complicated, and some

[00:20] guides certainly make it seem that way,

[00:22] but at its heart, it's surprisingly

[00:24] straightforward. Prompting is simply how

[00:26] we apply this course's description

[00:27] competency in practice. clearly

[00:30] communicating what we want, how we want

[00:31] it done, and how we want to interact

[00:33] with our AI assistant throughout the

[00:35] entire process. Think of prompting like

[00:37] explaining a task to a helpful new

[00:39] colleague who's eager to assist, but

[00:41] needs clear directions and expectation

[00:43] setting to do their best work. We'll be

[00:45] using Claude throughout this section,

[00:47] but these tips can be carried over to

[00:49] many other AI systems. You might have

[00:52] heard the term prompt engineering tossed

[00:54] around. Prompt engineering is simply the

[00:56] practice of designing effective

[00:58] instructions for AI systems like Claude.

[01:01] It's about crafting your questions and

[01:02] providing context in ways that help AI

[01:04] assistants understand exactly what you

[01:07] want. What's fascinating is that

[01:09] effective prompting blends familiar

[01:11] human communication skills with a few

[01:13] considerations specific to AI. Many

[01:16] principles that make for good human

[01:17] conversation, such as being clear,

[01:19] providing relevant context, and giving

[01:21] concrete examples, also apply when

[01:23] working with AI. Yet, there are

[01:25] differences, such as being more explicit

[01:27] about things humans could naturally

[01:28] infer, and accommodating the AI's

[01:31] limited context window, and sometimes,

[01:33] depending on the AI you're working with,

[01:35] using specific formatting that machines

[01:37] can easily process. As AI assistants

[01:40] continue to evolve, prompting best

[01:42] practices evolve, too. What works with

[01:44] today's AI systems may be different from

[01:46] what works with tomorrow's.

[01:48] Experimentation is key to discovering

[01:50] what works best for your specific needs.

[01:52] In this video, we'll mainly explore six

[01:54] foundational prompting tips that will go

[01:56] a long way toward helping you

[01:57] effectively communicate and collaborate

[01:59] with Claude and other AI systems. They

[02:01] are give Claude context, show examples

[02:04] of what good looks like, specify output

[02:07] constraints, break complex tasks into

[02:10] steps, ask Claude to think first, and

[02:13] define Claude's role, style, or tone.

[02:15] The first principle is simple but

[02:17] powerful. Be specific and clear about

[02:19] what you want, why you want it, and

[02:21] perhaps most surprisingly, who you are.

[02:24] Let's take a simple prompt. Tell me

[02:26] about climate change. How can we improve

[02:29] this by giving Cloud more context? A

[02:31] more specific contextrich version could

[02:33] look like, explain three major impacts

[02:36] of climate change on agriculture in

[02:38] tropical regions with examples from the

[02:40] past decade. Our baseline prompt was

[02:43] vague and leaves clawed guessing about

[02:44] our interests, level of knowledge, and

[02:46] the depth of detail we're looking for,

[02:48] such as geography and time span. We can

[02:50] even add more context by providing

[02:52] information not just about what we're

[02:54] looking for, but why we're asking and

[02:56] how we'll be using that information that

[02:58] Claude gives us. Now, our prompt looks

[03:00] like this. Explain three major impacts

[03:02] of climate change on agriculture in

[03:04] tropical regions with examples from the

[03:06] past decade. I'm preparing for a job

[03:09] interview at an agricultural research

[03:10] lab in Indonesia. I have a degree in

[03:13] ecology, but no specific knowledge on

[03:15] climate change. write a summary of key

[03:17] concepts that would help me speak

[03:18] intelligently in the interview. All this

[03:21] added context helps tailor Claude's

[03:23] response to your specific situation and

[03:25] knowledge level. This kind of background

[03:26] information is something we naturally

[03:28] provide in human conversations, but

[03:30] might forget to include when talking

[03:31] with Claude. Sometimes showing is better

[03:34] than telling. Providing examples of the

[03:36] kind of output you're looking for can be

[03:38] incredibly effective. This is sometimes

[03:40] called fshot prompting or nshot

[03:42] prompting in technical circles where n

[03:44] is the number of examples given but it's

[03:47] really just about showing the AI

[03:48] examples for it to emulate. For

[03:50] instance, take the following prompt.

[03:52] Please convert this technical statement

[03:54] to plain language. The platform

[03:56] implements end-to-end encryption

[03:57] protocols to safeguard data integrity.

[04:00] Clog may already be able to do this to

[04:02] your satisfaction. So we definitely

[04:04] recommend you just try first without

[04:06] examples and see where it leads you. But

[04:08] let's say you have a very specific style

[04:10] you want Claude to follow, and it's

[04:12] harder to explain than to give examples.

[04:15] Your refreshed prompt could look

[04:16] something like this. Here are two

[04:19] examples of how to convert technical

[04:20] jargon into plain language. Original,

[04:23] the quantum algorithm exhibits quadratic

[04:25] speed up. Plain, the new method solves

[04:27] problems roughly twice as fast as

[04:29] previous methods. Original, the

[04:32] interface leverages intuitive design

[04:33] paradigms. Plain, the design is easy to

[04:36] understand and use. Now, please convert

[04:39] this complex technical manual to plain

[04:41] language. When providing examples, aim

[04:43] to cover the full diversity of possible

[04:45] prompts, such as examples that cover

[04:47] different cases or styles. This helps

[04:50] Claude better understand the broad range

[04:51] of the pattern you want it to follow.

[04:53] Being clear about output constraints,

[04:55] such as the desired format and length of

[04:57] Claude's response, or the language you

[04:59] want Claude to code in, or the color of

[05:01] the buttons on the web page you want

[05:03] Claude to design, also helps ensure you

[05:05] get exactly what you need. Here's an

[05:08] example of clear and detailed

[05:09] description to ensure Claude delivers

[05:11] exactly what you're looking for. Create

[05:14] a clean, modern, single page art

[05:16] portfolio website. Include these main

[05:18] sections: hero, about me, skills,

[05:19] portfolio projects experience and

[05:21] contact. Make the navigation menu sticky

[05:24] and responsive with hamburger menu on

[05:25] mobile. Use a sunset color palette and

[05:28] add a dark light mode toggle in the

[05:30] navigation. Guidance like this helps

[05:32] cloud structure its response to match

[05:34] your expectations. When you have a

[05:36] complicated request, breaking it down

[05:38] into smaller steps helps Cloud follow

[05:40] your thinking and deliver better

[05:42] results. Think about it this way. If you

[05:45] ask a friend to do something for you

[05:46] without specifying how, there's a chance

[05:48] that they may not do it the way you

[05:50] intended them to. We've all been there.

[05:52] Listing out task steps ensures that

[05:54] Claude follows the process you want to

[05:56] in order to accomplish its task. This is

[05:59] sometimes called chain of thought

[06:00] prompting. For example, instead of

[06:02] asking Claude to analyze this quarterly

[06:04] sales data, you might say, "I'd like to

[06:06] analyze this quarterly sales data.

[06:08] Please approach this by looking through

[06:10] our sales records to identify the top

[06:11] performing products, comparing current

[06:14] quarter results to the previous quarter,

[06:16] highlighting any unusual trends or

[06:17] patterns, and then suggesting possible

[06:19] reasons for these trends. By default,

[06:21] you may not need to do this, especially

[06:23] for tasks that are relatively

[06:24] straightforward. Furthermore, modern

[06:27] reasoning models or extended thinking

[06:29] models are increasingly capable of

[06:31] performing step-by-step reasoning on

[06:33] their own, but you can still guide this

[06:36] process to ensure it aligns with your

[06:37] needs. The more variance there is in

[06:40] ways to execute the task well, or the

[06:42] more that proper task execution relies

[06:44] on experience and knowledge you've

[06:45] gained as a domain expert, the more you

[06:48] should consider taking the time to

[06:49] translate that knowledge into Claude.

[06:51] Relatedly, sometimes it can be helpful

[06:53] to explicitly give AI assistants like

[06:55] Claude space to work through its process

[06:57] first before executing its task. This

[07:00] approach helps Claude produce more

[07:02] thorough and well-considered responses.

[07:04] For example, you can add this to your

[07:06] prompt. Before answering, please think

[07:08] through this problem carefully. Consider

[07:11] the different factors involved,

[07:12] potential constraints, and various

[07:14] approaches before recommending the best

[07:16] solution. As I mentioned, modern

[07:18] reasoning or extended thinking models by

[07:20] default think before acting. But if

[07:22] you're working with an AI assistant that

[07:24] does not think first by default, you can

[07:26] still prompt the AI to do so. I want to

[07:28] note the importance of giving the AI

[07:30] assistant space to think before doing

[07:32] its task, not after. If you want that

[07:35] thinking to increase the quality of the

[07:37] AI's work, just like how having space to

[07:39] think before you act is different than

[07:41] acting first, then being asked to

[07:43] explain your thinking afterwards. As a

[07:45] side benefit, this also allows you to

[07:47] better see where the AI assistant might

[07:49] be going astray and thus where you could

[07:51] hone your description competency further

[07:53] by providing more guidance. Specifying

[07:56] how you want cloud to communicate and

[07:58] behave can significantly change how it

[08:00] approaches a task. By specifying the

[08:02] level of expected expertise, the

[08:04] perspective you want it to take, or its

[08:06] communication style, you can guide both

[08:08] Claude's interaction with you and the

[08:10] final result of what it produces. Simply

[08:12] put, who do you want the AI to act as?

[08:15] For example, take this prompt. Please

[08:17] explain how rainbows form from the

[08:19] perspective of an experienced science

[08:21] teacher speaking to a bright 10-year-old

[08:23] who's interested in science. This is

[08:25] also a good way to brainstorm or get

[08:27] feedback. You can specify a general role

[08:29] or even ask Claw to take on the persona

[08:31] of a specific figure, such as Richard

[08:32] Fineman, when asking for physics

[08:34] explanations. Here's another example. As

[08:37] a UX design expert, review this website

[08:40] wireframe and suggest three improvements

[08:42] focusing on user navigation and

[08:43] accessibility. Perhaps the most powerful

[08:45] technique is asking Claude to help

[08:47] improve your prompt. When you're not

[08:49] sure how to ask for something or how to

[08:51] improve your prompt, describe to Claude

[08:53] your issue or situation and ask it to

[08:55] make your prompt better or write your

[08:56] prompt for you. I'm trying to get you,

[08:59] Claude, to help me with goal. I'm not

[09:01] sure how to phrase my request to get the

[09:04] best results. Can you help me craft an

[09:06] effective prompt for this? Here's where

[09:08] Claude and other AI assistants may vary

[09:10] most in terms of performance. So, we

[09:12] suggest you experiment with different

[09:13] models as part of practicing delegation.

[09:17] Effective prompting is iterative and

[09:19] experimental. AI systems and best

[09:21] practices are constantly evolving. So,

[09:23] what works today may change tomorrow.

[09:25] Your first attempt won't always yield

[09:27] the perfect result, and that's expected.

[09:29] When a response isn't quite what you

[09:31] need, try refining your approach by

[09:33] playing around with any of the

[09:34] techniques we mentioned, such as add

[09:37] more specificity or context. Provide

[09:40] examples of your desired output. Break

[09:42] the task into smaller steps and try a

[09:45] different technique or combination of

[09:47] techniques. You can also ask for

[09:49] variations such as, "Can you give me

[09:50] three different versions of this?" You

[09:53] can request different formats such as,

[09:54] "Instead of a paragraph, could you

[09:56] present this in an interactive

[09:58] artifact?" Note that artifacts are a

[10:00] unique way that Claude can create

[10:01] outputs that may be easier to understand

[10:03] or more interesting to digest. You can

[10:05] also check confidence, such as for

[10:07] factual questions, you can ask, "How

[10:09] confident are you about this answer?"

[10:12] You can also reset the conversation

[10:13] entirely. Sometimes starting a fresh

[10:15] conversation gives better results than

[10:17] trying to correct the conversation

[10:19] that's gone off track. Use each

[10:21] interaction as feedback to improve your

[10:23] next prompt. Over time, you'll develop

[10:25] an intuition for how to communicate

[10:27] effectively with all AI systems. As you

[10:30] apply these techniques in practice,

[10:31] here's some guidance to recap. Some

[10:34] patterns consistently work well.

[10:36] Starting with a clear task overview

[10:37] statement, including format

[10:39] specifications and examples, setting

[10:41] explicit constraints or requirements,

[10:43] providing rich and relevant background

[10:45] information, and common mistakes to

[10:47] avoid are assuming that Claude can read

[10:49] your mind, or overloading a single

[10:52] prompt or conversation with multiple

[10:53] unrelated tasks, being too vague about

[10:56] what success looks like, and not

[10:57] providing feedback on previous

[10:59] responses. To recap, effective

[11:01] communication with AI systems like

[11:03] Claude combines timeless human

[11:04] communication principles with AI

[11:06] specific techniques. The approaches

[11:08] we've covered will serve you well across

[11:10] different AI systems. These six

[11:12] principles together with the secret

[11:13] weapon of asking Cloud for help form a

[11:16] solid toolkit for applying the

[11:17] description competence to your AI

[11:19] interactions. Iteration and practice

[11:21] here is the key to Swift improvement and

[11:24] mastery. Remember that prompt

[11:26] engineering is an evolving practice. As

[11:28] models improve, some specific techniques

[11:30] become less necessary. However, these

[11:33] principles of good communication are

[11:35] still relevant even if the way we apply

[11:37] them changes. Maintain a spirit of

[11:39] experimentation and adapt your approach

[11:42] based on your results.