Week 10: Advanced Prompting and Optimizing Results

Learning objectives

By the end of this week, you will be able to:

1. Explain what a system prompt or custom instructions is, how it differs from a single message, and how it connects to a feature you've already used in an AI chat tool that lets you set standing instructions.
2. Ask an AI tool to reason step by step before giving a final answer, a technique called chain-of-thought prompting, and explain why this can improve accuracy on harder, multi-step questions.
3. Explain the caveat that reasoning text produced by an AI tool can still contain hallucinations and should still be checked, using verification habits you have already been practicing.
4. Build a few-shot prompt that gives an AI tool two or three examples of the input and output pattern you want, and explain why more examples tend to produce more consistent results.
5. Break a large writing or summarizing task into smaller prompted steps, such as requesting an outline first, instead of asking for the entire finished result in one prompt.
6. Compare the outputs of two different AI tools on the same well-formed prompt, using a second-opinion verification technique you have already practiced.
7. Start a personal, reusable prompt library, saved as a plain text or Markdown file, that you will continue building for the rest of this course.
8. Combine several advanced prompting techniques within a single realistic task.

Lesson

From single messages to advanced technique

The prompt-writing fundamentals you have already practiced, giving clear instructions, providing context, offering examples, assigning a role to the AI tool, requesting a specific output format, iterating when the first response is not quite right, and prompting specifically for accessible output, such as asking for plain text and structured headings instead of tables or dense visual layouts, focus on one message at a time: you write a clear, well-formed prompt, send it, read the response with your screen reader, and revise if needed. That approach still works, and you should keep using it. This lesson adds a second layer on top of it. Instead of only improving what you say in a single message, you will learn ways to shape an entire conversation, guide an AI tool's reasoning process, teach it a pattern through multiple examples, split a big job into a sequence of smaller prompts, and check a result by asking a second AI tool the same question. Together, these techniques are what separate an occasional AI user from someone who reliably gets strong results on harder, more valuable tasks, exactly the kind of prompt fluency this course has identified as foundational to almost everything else you do with AI.

None of these techniques require anything beyond what you already have: a web browser or app for an AI tool such as Claude or ChatGPT, your screen reader, and the typing and navigation skills you have been building throughout this course. What changes in this lesson is not the tools, but how deliberately and strategically you use them.

System prompts and custom instructions: standing instructions instead of one message

A system prompt is a standing instruction that applies to a whole conversation, or to every conversation you start with a particular AI tool, rather than to just one message. Think of the difference between telling a coworker something once, for a single task, versus giving them a general instruction they will remember and follow every single time they help you from then on, such as "always send me a plain-text summary before the full document." A system prompt is that second kind of instruction, aimed at an AI tool instead of a coworker.

You have already met a version of this idea. If you have used ChatGPT, you may recognize its custom instructions feature: a settings screen with two boxes, one where you describe yourself and your situation, and one where you describe how you want ChatGPT to respond. Once you fill in those boxes, ChatGPT applies that guidance automatically to every new conversation, without you needing to repeat it. That is a system prompt in practical, everyday form. Claude offers a similar standing-instruction layer through Projects: instructions you set once for a Project apply to every conversation inside that Project. Some AI tools also let a developer set an underlying system prompt that shapes the AI's behavior before you ever type a word, which is part of why the same AI model can feel a little different depending on which app or website you are using it through.

Custom instructions are especially useful for exactly the kind of standing needs a blind or low vision learner is likely to have. Instead of retyping "please use plain text with clear headings, and do not put important information only in a table" at the start of every single conversation, you can set that preference once as a custom instruction and let it apply automatically going forward. Other useful standing instructions include a preferred reply length, a preferred tone, a note that you are a screen reader user so any described visual layout should be explained in words, or a reminder to flag when the AI tool is uncertain about a fact rather than stating it as if it were definitely true.

It helps to keep the levels straight. A single message is the most specific and temporary layer: it applies only to the one request you just typed. Custom instructions, set in your account or app settings, are a personal standing layer: they apply to every conversation you start, until you change them again. A system prompt set by the developer of an app or tool is the broadest and most fixed layer: it applies to everyone using that particular product, and you usually cannot see or change it directly, though you can often work around or add to it with your own custom instructions. Understanding these three layers helps explain why an AI tool sometimes behaves in a way you did not directly ask for in your message: it may be following a standing instruction you set earlier, or one the tool's developer built in.

This lesson gives you a working understanding of standing instructions so you can start using ChatGPT's custom instructions or Claude's Project instructions more deliberately right away. A later lesson in this course will return to this idea in much greater depth, helping you build a complete, personal, accessibility-first instruction set that you can reuse across tools and tasks. Treat what you learn here as the foundation that later lesson will build on, not the finished result.

Chain-of-thought requests: asking an AI tool to think step by step

Chain-of-thought prompting means explicitly asking an AI tool to reason through a problem step by step, or to show its reasoning, before giving you a final answer. In practice, this can be as simple as adding a sentence to your prompt, such as "think through this step by step before you answer" or "show your reasoning, then give me a clear final answer." Some AI tools also have a "thinking" or "reasoning" mode you can turn on, which does something similar automatically.

To understand why this technique tends to improve accuracy, it helps to recall a piece of vocabulary from earlier in this course: a large language model generates its response one token at a time, and each new token is generated based on everything that came before it in the conversation, a stretch of text called the context window. If you ask a hard, multi-step question and demand only a short final answer, the AI tool has to jump straight to a conclusion with nothing but your original question in its context window to draw from. If you instead ask it to reason out loud first, each step of that reasoning becomes part of the context window before the final answer is generated. The model is, in effect, building up more relevant, specific text to draw its next tokens from, rather than jumping straight from a short question to a guess. This tends to produce more accurate results on questions that involve several steps, such as a multi-part math problem, a logical puzzle, a scheduling conflict with several constraints, or a decision that depends on weighing more than one factor at once.

Chain-of-thought prompting is not equally useful for every task. For a simple factual question with one clear answer, asking an AI tool to "think step by step" mostly adds extra text to read without much benefit. It earns its keep on genuinely multi-step problems: working out a budget across several expenses, planning a sequence of job application steps with overlapping deadlines, debugging why a set of instructions did not produce the expected result, or working through a multi-part comparison between two options. As a rule of thumb, if a problem would take you more than one clear step to solve by hand, chain-of-thought prompting is worth trying.

Here is an honest and important caveat, directly connected to an earlier lesson on AI's weaknesses: the reasoning text an AI tool produces can still contain hallucinations, meaning confidently stated information that is wrong. A chain-of-thought response is not a guarantee of correctness. It is entirely possible for an AI tool to write out several steps that sound logical and organized, include one subtly wrong fact or calculation partway through, and then confidently deliver a wrong final answer built on that mistake. Reading the reasoning is actually one of the most useful ways to verify a result, since it lets you check each step rather than only the final claim, but it does not replace verification. Apply the verification habits you have already learned: check any factual claim that matters against an independent source, and treat a confident tone as no guarantee of accuracy, whether that confidence appears in a short answer or in several paragraphs of reasoning.

One accessibility-specific tip is worth adding here. Long chain-of-thought responses can be tiring to read in full with a screen reader, especially if the reasoning wanders. You can combine this technique with the output-format requests you have already learned by asking for something like: "Show your reasoning briefly, in a few short steps, then give me a clear final answer under its own heading." This keeps the accuracy benefit of chain-of-thought reasoning while keeping the response itself easy to navigate by heading, which matters as much for a technique's usefulness to you as its raw accuracy does.

Few-shot examples: showing the pattern instead of just describing it

You have already learned to include an example in a prompt when it helps the AI tool understand exactly what you want. This lesson extends that idea: a few-shot example means giving an AI tool two or three examples of the input-and-output pattern you want, together in a single prompt, rather than only one. The term comes from the idea of the AI tool learning a pattern from a "few shots," or examples, before applying it to a new case of your own.

Consider a concrete illustration. Suppose you want an AI tool to turn messy meeting notes into clean, short task items in a consistent format. A single example might look like this: "Note: need to email Sarah about the budget by Friday. Task: Email Sarah about the budget. Due: Friday." That one example shows the AI tool roughly what you want, but it leaves real questions unanswered: should every task include a due date, even when the note does not mention one? Should the task line always start with a verb? What happens with a note that describes two tasks at once? A few-shot prompt answers these questions by example instead of by additional instructions. You might provide three short example pairs: one with a clear deadline, one with no deadline at all, and one where the original note describes two separate tasks that need to become two separate task lines. Seeing all three examples together, the AI tool has a much clearer picture of the exact pattern you want, including how to handle the tricky cases, and it is far more likely to apply that same pattern consistently to new notes you give it afterward.

This is the core reason few-shot examples tend to produce more consistent results than a single example or a purely descriptive instruction: examples show edge cases, formatting choices, tone, and length all at once, in a way that is hard to fully describe in words alone, especially in a short prompt. A single example can be misread as showing one narrow case rather than a general rule. Multiple, varied examples make the general rule much clearer.

When you build a few-shot prompt, keep the examples short, keep their format consistent with each other, and choose examples that represent the real variety you expect to encounter, including at least one slightly unusual or tricky case if you can think of one. Three well-chosen examples are usually more valuable than five repetitive ones that all show the same simple pattern. This technique pairs naturally with the personal prompt library you will start building later this week, since a few-shot prompt that works well for a recurring task, such as cleaning up notes, formatting a weekly report, or converting a document into accessible plain text, is exactly the kind of reusable prompt worth saving rather than rebuilding from scratch every time.

Breaking a big task into smaller steps

A single giant prompt asking for an entire finished report, a full multi-page document, or a summary of a very long file all at once often produces a weaker result than the same task broken into a sequence of smaller, more focused prompts. There are two related reasons for this. First, a very large request gives the AI tool a lot of ground to cover at once, which tends to produce a more generic, less carefully considered result section by section, the same way a person asked to write an entire report in one uninterrupted sitting tends to produce weaker work than one who plans it out first. Second, and just as important for you specifically, a giant finished result is much harder to verify. Reading and checking one long document for accuracy, tone, and completeness all at once, especially with a screen reader, is far more tiring and far more likely to let an error slip past you than reviewing several shorter pieces as you go.

The most useful version of this technique is asking for an outline first. Instead of "write me a full report on our department's quarterly progress," try "give me a short outline of a quarterly progress report, with section headings only, no full paragraphs yet." Read the outline, adjust it, add or remove sections, and only once you are satisfied, ask the AI tool to fill in one section at a time: "Now write the first section, on completed projects, in full." This turns one large, hard-to-verify task into a short planning step followed by several small, easy-to-verify steps, and it gives you a natural checkpoint to catch a wrong direction early, before a great deal of text has already been written around it.

A second version of this technique applies to summarizing a long document: instead of pasting an entire lengthy file into one prompt and asking for a single summary, break the document into a few logical chunks, such as by section or by a fixed number of pages, and ask the AI tool to summarize each chunk on its own. Once you have a short summary of each chunk, you can either read them individually or ask the AI tool to combine those chunk summaries into one final, shorter overview. This chunked approach tends to preserve more detail than a single request to summarize a very long document all at once, since the AI tool is working with a smaller, more focused piece of context window each time rather than trying to compress an enormous amount of text into a few sentences in one pass. You will get direct, hands-on practice with reading and working through long files a section at a time in a later part of this course, when it turns to working files such as Markdown documents directly.

This step-by-step approach takes a little more of your own time than firing off one giant prompt and hoping for the best, but it consistently produces stronger, more accurate, and far easier to verify results, which makes it well worth that extra time for any task that genuinely matters, such as a resume, a work report, or a document you plan to share with someone else.

Comparing outputs across models

You have already learned to get a second opinion by checking an AI tool's claim against an independent source rather than accepting it on faith. This lesson extends that same idea into a specific, practical prompting technique: deliberately sending the exact same, well-formed prompt to two different AI tools, such as Claude, ChatGPT, or Gemini, and comparing the two answers side by side.

This technique serves two distinct purposes. The first is verification. If two different AI tools, built by different companies and trained somewhat differently, independently produce the same or a very similar answer to a factual or analytical question, that agreement gives you more confidence in the result than either answer alone. If the two tools disagree, that disagreement is a clear, concrete signal to slow down and check the claim against an outside, independent source before relying on it, exactly the kind of verification habit you have already practiced as essential to using AI responsibly. Comparing across models does not replace outside verification for anything important, but it is a fast, easy first check that often catches a problem before you go looking any further.

The second purpose is learning each tool's particular strengths, so you know which one to reach for on a given kind of task in the future. Over time, as you use this comparison technique on different types of requests, you may notice patterns: perhaps one tool tends to give longer, more thorough answers while another tends to be more concise, or one handles a long document you uploaded through a feature like Claude's Projects more smoothly, while another has a voice mode that fits a particular workflow better. These patterns are worth noticing and remembering, since they let you pick the right tool for a given job going forward instead of always defaulting to whichever one you happen to open first.

To use this technique with a screen reader, keep the prompt itself identical, or as close to identical as the two tools' interfaces allow, since even small wording differences can produce different results and make the comparison less fair. Open each tool, either in two separate browser tabs you switch between with your screen reader's tab-switching commands, or one after another, and consider copying both responses into a single plain-text document with a clear heading for each tool's answer, so you can read and compare them together using your screen reader's heading navigation, rather than having to hold both responses in memory as you switch back and forth between windows.

Building a personal prompt library

A prompt library is a personal, saved collection of prompts that worked well for you, kept in a plain text or Markdown file, organized so you can find and reuse them instead of reinventing a good prompt from scratch every time you need it. Once you have written a prompt that reliably produces a good result, whether it is a few-shot prompt for cleaning up notes, a chain-of-thought prompt for working through a scheduling problem, or simply a clearly worded request that took you a few tries to get right, saving it means that effort pays off every future time you need something similar, instead of only once.

A later lesson in this course will teach you the details of working with plain text and Markdown files directly, including how to read and edit them with your screen reader in Notepad and VS Code, so for now you only need a simple starting version: a single plain text file, saved somewhere you can find it again, such as your portfolio folder. A useful structure for each entry in your prompt library has three parts: a short task name that describes what the prompt is for, the prompt text itself exactly as you would type it, and a brief note about what worked well or what you would adjust next time.

Here is what one entry might look like in practice:

Task: Clean up messy meeting notes. Prompt: "Turn these notes into a short numbered task list. Each task should start with a verb and include a due date if one is mentioned, or say 'no deadline noted' if not. Here are three examples of the format I want: [examples]. Now do the same with these notes: [notes]." Note: works well with two or three examples included; without examples, the AI tool sometimes invents due dates that were not in the original notes.

A prompt library does not need to be long to be useful, and it is meant to grow gradually rather than all at once. A handful of solid, reusable prompts, each with an honest note about what worked and what to watch for, is more valuable than a long, disorganized list of prompts you never look back at. This week's portfolio project asks you to start exactly this file, and you will keep adding to it for the rest of the course. It will become directly useful again later in this course, when you build a fuller personal instruction set and learn to build simple agents that can reuse a saved prompt automatically as part of a repeating task.

Putting it together: a worked example

These techniques are most powerful combined, applied together to one real task rather than used one at a time in isolation. Here is a short, realistic example that pulls together most of what this week has covered.

Imagine you are preparing to summarize a long article for a coworker who does not have time to read the whole thing. First, you set a standing custom instruction in your AI tool, established once rather than repeated every time, asking it to always reply in plain text with short paragraphs and no tables, matching the accessible-output habit you have already practiced. Next, instead of pasting the entire article into one prompt, you break it into three sections and ask for a short summary of each section on its own, applying this lesson's step-by-step technique. For the most complicated section, the one with several statistics and comparisons, you add a chain-of-thought request, asking the AI tool to first list the key numbers it found, then reason about what they mean together, before giving you a two-sentence summary, since this is exactly the kind of multi-step task where that technique earns its keep. You give the AI tool two short few-shot examples of the summary style you want, so all three section summaries come out in a consistent tone and length. Once you have all three section summaries, you send the same final combined-summary prompt to a second AI tool as a quick check, comparing the two results before you send anything to your coworker. Finally, because this combination of techniques worked well and you expect to summarize long articles again, you save the whole prompt sequence, with a short note about what worked, in your prompt library.

Notice that no single step in that example was complicated on its own. Each technique, a standing instruction, a broken-down task, a chain-of-thought request on the hardest part, a few-shot example for consistency, a second opinion from another tool, and a saved record for next time, is simple by itself. Combining them deliberately, on a task that actually matters to you, is what turns prompt fluency from a helpful habit into a genuinely professional skill.

Key terms from this week

Use this list to review the vocabulary introduced in this lesson before you start the exercises and the test.

System prompt

A standing instruction that applies to a whole conversation, or to every conversation with an AI tool, rather than to just one message.

Chain-of-thought

A prompting technique that explicitly asks an AI tool to reason through a problem step by step, or show its reasoning, before giving a final answer.

Few-shot example

A prompt that includes two or three examples of the input-and-output pattern you want, rather than only one example.

Prompt library

A personal, saved collection of prompts that worked well, kept in a plain text or Markdown file so they can be reused and adapted instead of rewritten from scratch.

Keyboard-only exercises

These exercises use only your keyboard. Screen reader commands are described in general terms below; substitute your own screen reader's equivalent command where needed.

Exercise 1: Compare a direct answer with a chain-of-thought answer

1. Open an AI tool you are comfortable with, such as Claude or ChatGPT, and start a new conversation.
2. Move focus to the message box. Press Tab repeatedly, or use your screen reader's browse-mode command for jumping to edit fields or form controls, until your screen reader announces the message edit field. Check your own screen reader's shortcuts if you are unsure which key moves focus directly to a form field.
3. Type a moderately hard question that involves more than one step, such as a multi-part scheduling problem or a question that requires weighing two or three factors together, and press Enter to send it.
4. Read the response with Say All, or your screen reader's continuous reading command.
5. Return to the message box and send the same question again, this time adding "Think through this step by step, then give me a clear final answer" to the end of your prompt.
6. Read the second response the same way, and compare it to the first: did the step-by-step version reach a different, more accurate, or better-explained answer?

Exercise 2: Compare the same prompt across two AI tools

1. Write one well-formed prompt, following the clear-instruction habits you have already practiced, on a topic you are genuinely curious about.
2. Send that exact prompt to one AI tool, such as Claude, and read the response with Say All.
3. Open a second AI tool, such as ChatGPT or Gemini, in a new browser tab. On most browsers, Ctrl plus T opens a new tab, and Ctrl plus Tab switches between open tabs. On a Mac, use Command plus T and Command plus the number of the tab, or your screen reader's tab navigation commands.
4. Send the exact same prompt, word for word, to the second tool and read that response with Say All as well.
5. Compare the two answers. Note whether they agree, where they differ, and which response you found clearer or more useful, applying the second-opinion habit you have already learned.

Exercise 3: Start your prompt library file

1. Open Notepad. On Windows, you can search for it from the Start menu by typing "Notepad" and pressing Enter on the matching result.
2. Type a title at the top of the file: "My Prompt Library."
3. Choose one prompt from Exercise 1 or Exercise 2 that worked especially well, and type it into the file, labeled with a short task name, such as "Task: Comparing scheduling options."
4. On the line underneath, write a one-sentence note about what worked well or what you would change next time you use it.
5. Save the file with the name my-prompt-library, in the same portfolio folder you have used throughout this course, using Ctrl plus S and typing the file name when prompted.

Portfolio project: My Prompt Library

This week's portfolio piece is the real, ongoing start of your personal prompt library, the same file you began in Exercise 3. You will keep adding to this file for the rest of the course, so treat it as a living document rather than a one-time assignment.

1. Open the my-prompt-library file you created in Exercise 3, or open Notepad and start a new file with that name if you have not already.
2. Add at least five prompts you have found genuinely useful across the AI tools you have learned so far, drawing on prompts you wrote while learning Claude, ChatGPT, Gemini, or any other tool covered so far, as well as the prompts from this week's exercises.
3. For each of the five prompts, include a short label describing the task it is for, the prompt text itself, and one sentence explaining what it is good for or when to reach for it.
4. Reread all five entries with Say All, and check that each one is specific enough that you could copy and reuse it later without having to remember extra context.
5. Save the file, keeping the name my-prompt-library, in your portfolio folder.
6. Add a short note at the top of the file reminding yourself that this is an ongoing collection: plan to keep adding useful prompts to it every week for the rest of the course.

Weekly test

Answer key

Each answer below includes a one-sentence explanation drawn from this week's lesson.