I wanted to get better at prompt engineering. Not the trick-the-robot kind, the boring-but-useful kind: how to ask a model a question so you get an answer you can actually trust.

The trouble with practicing is that most tutorials use made-up examples, and it's hard to tell a good answer from a bad one when you don't care about the topic. So I practiced on something I did care about: the deals sitting in my Amazon cart. I had a vacuum I'd been eyeing and a hair styler that was "43% off," and I genuinely wanted to know if those were good prices or just good marketing. The stakes were real, actual money on an actual decision, and that's what made it a good drill. A vague prompt gives you a confident answer, and when you actually care, you can feel that the answer is hollow.

What I learned, with the real deals and the actual before-and-after prompts:

The trap hiding in every deal

Start with the hair styler. The listing said:

Shark FlexStyle. Limited time deal. $199.00, 43% savings. List Price: $349.99.

My first instinct was the prompt most people write:

"Shark FlexStyle $199, 43% off list $349.99, is that a good deal?"

This feels reasonable. It is also nearly useless: it lets the model answer the easy question (is 43% off a big discount? sure!) instead of the real one (is $199 actually a good price?). That $349.99 list price is a marketing anchor. A lazy prompt accepts it, and so you get a lazy "yes, great deal!" back.

The fix was re-framing this:

Act as a pricing analyst. I don't care whether $199 looks like a discount off list. I care whether $199 is a genuinely good price for the Shark FlexStyle right now. Before concluding, work through: (1) the actual street price over the last 6-12 months, (2) how often it drops to or below $199, (3) the real discount vs. its typical selling price, not vs. list. Cite a source and date for each price, or mark it unverified.

Same question, completely different answer. What the assistant came back with, in its own telling:

$199 is a genuinely good price, but not for the reason Shark markets it. It's the 6-month floor and has only appeared a couple of times in that window. The more honest benchmark is its recent median price, and $199 is a real savings off that, not the headline savings off the inflated $349.99 list.

I'm taking those specifics on faith, the assistant reported them and I didn't independently verify its price history, but that's almost beside the point. What changed was the shape of the answer: instead of "43% off, great deal!" I got "$199 is the all-time low on Amazon - only hit twice in 180 days, and the previous touch was one day only" The only thing I did differently was tell the model to ignore the anchor and compare to street price.

Lesson one: reframe the question to attack the anchor. "Compare to street price, not list" is the line I am going to find more helpful going forward. It applies to nearly every "is this a good deal" question you'll ever ask.

When "better" has no meaning yet

Next I tried a comparison, which is where vague prompts fall apart:

Dyson V8 Cordless: $269.99, 31% off list $389.99.

Shark IZ363HT with HEPA filter: $199.99, 43% off list $349.99.

Just the plan old compare with similar on Alexa for Shopping on Amazon

The trap is right there in the numbers. The Shark is 43% off and cheaper, so a lazy prompt crowns it the winner in one line. But "better deal" is meaningless until you say better for what. A comparison has no answer until you define the criteria.

The prompt I landed on added three new moves on top of the anchor break:

Act as a product and pricing analyst. Don't let a bigger discount percentage decide the winner. Compare the Dyson V8 ($269.99) and Shark IZ363HT ($199.99) as a purchase. Ignore list-price discounts; compare each to its real street price. Flag if they're different product classes. Score on weighted criteria: price-vs-street 30%, cleaning 25%, runtime 20%, HEPA 15%, upkeep 10%. Then answer two things separately: better DEAL, and better PRODUCT. Cite source and date per price or mark unverified.

Three things made this work:

1. Weighted criteria. "Better" means nothing until you declare what matters and how much. Writing the weights forced me to admit what I actually cared about.

2. Split the deal from the product. A thing can be a great deal and the wrong buy. Keeping those two questions separate clears up most of the confusion in a comparison.

3. Flag the class mismatch. These two vacuums aren't the same tier, and I wanted the model to say so instead of comparing raw prices.

The result split exactly the way good analysis should (again, these are the assistant's figures, which I didn't independently confirm):

Better DEAL: the Shark at $199.99, a genuine discount off its typical street price. The Dyson's deal is real but based on thin historical data. Better PRODUCT: the Dyson V8, for cleaner engineering and easier maintenance, though the Shark's HEPA seal is a meaningful win for a cat household.

One prompt, two different winners. A single "which is better?" would have collapsed that into one arbitrary answer and I'd have learned nothing.

Lesson two: in any comparison, the quality of the answer is capped by whether you defined the criteria. The model can't tell you "better" until you tell it "better how."

The model is smarter, not infallible

The part that surprised me most came in that same comparison answer, which also contained this:

The $130 gap vs. the Dyson is real money…

and later:

…are comfortable paying $70 more.

Both can't be true. The actual gap is $269.99 minus $199.99, which is $70. The $130 was wrong, and the model contradicted itself within the same answer. A well-engineered prompt made the reasoning much better. It did not make the arithmetic correct.

The fix is one more constraint: "show the price-difference math explicitly." Errors hide in summary claims but not in shown work. If I'd forced it to write $269.99 - $199.99 = $70, the contradiction would have been impossible.

Lesson three: a better prompt improves reasoning, not truth. Always verify the model's math.

The constraint I should've expected: word count

When I copied that comparison prompt into the Amazon shopping assistant, only the first half pasted. There was a character limit, and my nicely-formatted prompt with its ROLE and TASK and OUTPUT headers got cut off mid-sentence.

The annoying part is that nothing told me the limit existed until I hit it. These assistants rarely document it, so you find the ceiling by bumping your head on it. That changes the strategy: since you can't plan around a number you don't know, put the instructions that change the answer first, so if you do get truncated, you lose the formatting and not the substance.

That accident taught me something useful. Under a tight budget, you rank techniques by how much each word buys you and cut the rest. The headers and scaffolding? First to go, the model infers structure on its own. What you protect, in order:

1. The anchor break (street price, not list)
2. The deal-vs-product split
3. The weighted criteria (compressed to inline shorthand)
4. The cite-or-unverified rule

The compressed version was a quarter the length and carried every essential instruction:

Compare as a purchase: Dyson V8 $269.99 vs Shark IZ363HT (HEPA) $199.99. Ignore list-price discounts; compare each to its real street price. Note if they're different product classes. Score on: price-vs-street 30%, cleaning 25%, runtime 20%, HEPA 15%, upkeep 10%. Answer two things separately: better DEAL, and better PRODUCT. Cite source+date per price or mark unverified.

Now with more details that you care about in the prompt. The engineered prompt, with the price-history table and chart it produced.

Lesson four: a prompt's power lives in its constraints and reframes, not its formatting. Headers are a readability luxury. The instructions that change the answer are the ones you never cut.

The same question, side by side

This is the comparison I'd point a skeptic to. I ran a third product through both kinds of prompt so you can see the difference in one frame: the Garmin epix Pro (Gen 2) Sapphire Edition, 51mm. A premium watch with a real list price north of $1,000, which makes the anchor problem even more tempting.

Without prompt engineering, you write what feels natural:

"Is the Garmin epix Pro (Gen 2) Sapphire 51mm a good deal right now?"

And you get back something that sounds authoritative and tells some but maybe not all you want:

The vague prompt, the vague(ish) answer.

Notice what that answer does: it accepts the discount at face value, and uses list price.

With prompt engineering, you spend the reframes:

Act as a pricing analyst. I care whether the current price of the Garmin epix Pro (Gen 2) Sapphire 51mm is genuinely good, not whether it looks discounted off list. Before concluding: (1) find its street/historical price over the last 6-12 months, (2) note how often it hits the current price or lower, (3) state the real discount vs. typical selling price, not list. Cite source and date per price or mark unverified. Then give a verdict: buy / wait / skip, and a "wait-for" target price if I should hold out.

The shape of what comes back changes entirely. Instead of a sales pitch, you get a verdict with the reasoning shown:

With some real, actionable information and price history in context:

That is a decision, with the math shown and the uncertainty labeled. One answer is a brochure, the other tells you exactly how good the price is, how often it appears, and what to hold out for. The difference isn't the model. It's the four or five sentences of reframing you put in front of it.

And one last thing, the part I didn't expect. The first time I ran this, the assistant answered from only about 40 days of data and called it a ~$50 saving. The phrase "over the last 6-12 months" in my prompt is what made it go back, pull a full year of price history, and correct itself to the ~$75 figure with a proper wait-for target. The reframe didn't just change the format of the answer. It changed how hard the model looked.

The cheat-sheet

The whole thing distilled, for when you want to try this yourself.

The five core techniques

1. Role for clarity, not theater. Set a role only when it changes what good output looks like, and say why. "Act as a pricing analyst; I care about real price history over marketing claims."

2. Explicit output structure. Tell it the shape: a table, then a verdict, then a confidence rating. Structure is the highest-leverage habit.

3. Few-shot examples. Show one example of the answer you want. It pins down tone and granularity better than any description.

4. Reason before verdict. Ask for the work first, conclusion last. Evidence-first ordering kills confident-but-wrong answers.

5. Constraints, especially for uncertainty. "If you can't verify it, say unverified rather than guessing." The uncertainty clause is the one people skip, and it's what prevents fabrication.

For comparisons, add

• Define and weight your criteria. No criteria, no real answer.
• Split "best deal" from "best product." Don't let a steeper discount pick the winner.
• Force the math to be shown so errors can't hide in summaries.

For deals specifically

• Break the anchor: compare to street price, never to list price.
• Ask how often the price appears, not just how low it is.

Meta-habits

• Specificity beats politeness. The model can't read intent you didn't state.
• Prompt engineering is debugging, not authorship. Run it, find the specific failure, fix that one thing.
• Watch which constraints the model quietly drops, then promote those to non-negotiable on the next pass.
• Under a word limit, cut formatting first and reframes last.
• Name the scope. A phrase like "over the last 6-12 months" makes the model look harder, not just format differently; it pulled a full year of data and corrected its own answer.

Reading the patterns vs. using them

None of this is new, and I knew that. I'd read White et al. (2023), A Prompt Pattern Catalog to Enhance Prompt Engineering with ChatGPT a while back. It catalogs prompt techniques as reusable "patterns," the way software design patterns work, and the moves I leaned on map almost one to one onto it:

• "Act as a pricing analyst" is the Persona Pattern: give the model a point of view so it knows what to focus on.
• "Compare to street price, not list" is the Question Refinement Pattern: get the model to answer a better version of the question than the one you asked.
• "Work through (1), (2), (3) before concluding" is the Cognitive Verifier Pattern: break a question into sub-questions and build the answer from them.
• "Cite each price or mark it unverified" and "verify the math" are the Fact Check List and Reflection patterns: make the model surface what its answer depends on, then check itself.
• The assistant offering to "pull price history on any specific item before you commit" is the Flipped Interaction Pattern, where the model drives by asking you questions.

But knowing the catalog and using it are different skills. Reading the paper, the patterns felt like a tidy list to nod along to. It wasn't until I was staring at a $549 watch I actually wanted, with real money on the line, that "Question Refinement" stopped being a term and became a reflex: no, don't accept the list price, make it compare to street. The patterns only stuck once I had a reason to care whether the answer was right.

If you want an easier intro than the paper, Descript's 5 Advanced Prompts to Get Better Answers from ChatGPT walks through five of these same patterns with plain examples. Read either one. Then go find a question you actually care about and use them, because that's the part the reading can't do for you.

The actual takeaway

I set out to learn prompt engineering and ended up also learning that the Shark IZ363HT was a genuinely good deal and the Dyson was the better vacuum, which is its own kind of useful. Still don't know which to buy, but at least I do know more and have developed some better prompting habits. Every one of these techniques is just a way of refusing to accept the easy answer: don't trust the list price, don't accept an undefined "better," don't believe the math without seeing it.

Pick something you actually want to know the answer to. That's the trick. When you care, you can feel the difference between a hollow answer and a real one, and chasing that difference is the entire skill.