What Prime Day Taught Me About Prompt Engineering
I wanted to master prompt engineering.
I did not want fancy tricks. I wanted useful skills. I wanted to know how to ask questions so the answers are trustworthy.
Most tutorials use fake examples. It is hard to spot a bad answer if you do not care about the topic.
So I practiced on my Amazon cart. I wanted to know if my deals were real or just marketing. The stakes were my own money.
Here is what I learned.
- Attack the Anchor Marketing uses "list prices" to make discounts look huge. If you ask "Is this a good deal?", the model often accepts the fake list price.
The fix: Tell the model to ignore the list price. Ask it to compare the current price to the real street price over the last 6-12 months.
- Define your criteria A prompt like "which is better?" is useless. "Better" means nothing without context.
The fix: Use weighted criteria. Tell the model exactly what matters.
- 30% price
- 25% cleaning power
- 20% runtime
- 15% HEPA filter
- 10% upkeep
Also, split "best deal" from "best product." A product can be high quality but a bad value.
- Force the math Models can be confident but wrong about math. They often make errors in summaries.
The fix: Tell the model to "show the math explicitly." If the model writes out the subtraction step by step, it is less likely to hallucinate the result.
- Prioritize substance over formatting If you hit a character limit, you lose the end of your prompt.
The fix: Put your most important instructions first. Do not waste space on long headers or polite words. Put the "anchor break" and the "weighted criteria" at the top.
5 Core Techniques to Use: • Use roles for clarity. Tell it to act as a pricing analyst. • Demand a specific structure. Ask for a table or a specific verdict. • Use few-shot examples. Show it one example of a perfect answer. • Ask for reasoning before the verdict. This forces the model to think. • Add an uncertainty clause. Tell it to say "unverified" if it cannot find data.
Prompt engineering is debugging. Do not just accept the first answer. Find where the model fails and fix that specific part.
Source: https://dev.to/cseeman/what-prime-day-taught-me-about-prompt-engineering-3gek
Optional learning community: https://t.me/GyaanSetuAi