Lesson 5 of 8 · 10 min read
“Prompt engineering” is an unfortunate name. It suggests something technical — a discipline with syntax you have to learn, tricks you have to memorise, a certification you have to buy.
It isn’t. It’s the practice of getting better answers by working with the thing rather than firing one question at it and walking away.
Lesson 4 covered the single good prompt. This lesson covers what happens after — which is where most of the value actually is, and where most beginners quit.
The first answer is a draft
The single biggest shift: stop treating the first response as the answer.
Beginners send a prompt, read the reply, decide it’s mediocre, and conclude the tool is overrated. Experienced users send a prompt expecting the first reply to be roughly 60% right — and then get to work.
That’s not a workaround for a flaw. It’s how the tool is meant to be used. You’re not querying a database. You’re in a conversation where you can say “no, not like that,” and it costs you eight seconds.
The people getting real value out of AI aren’t writing better first prompts. They’re on their fourth message.
Feedback that actually works
But how you push back decides everything.
Vague feedback gets vague changes:
❌ “Make it better.”
❌ “I don’t like this.”
❌ “Try again.”
The model has no idea what you objected to. So it reshuffles at random and you get a different flavour of the same problem.
Specific feedback works immediately:
✅ “Too formal. Drop the opening pleasantries and get to the point in the first line.”
✅ “Second paragraph is good, keep it. Rewrite the rest to match that tone.”
✅ “You’ve assumed they already know about the delay. They don’t. Start from there.”
Same instinct as Lesson 4: be precise about the gap between what you got and what you wanted. Naming what’s wrong is most of the work — and the useful side effect is that you’re forced to know what you actually want, which you often didn’t.
Keep what’s working. “The structure’s right, the tone’s wrong” saves the parts you liked. Without that, you’ll fix one thing and break another, and go in circles.
Giving it a role — and why it’s oversold
You’ll see advice everywhere to start prompts with “Act as a world-class marketing expert with 20 years of experience…”
Be sceptical. This has become cargo cult — repeated because it’s repeated, rarely tested. Piling on superlatives does very little. “World-class” and “expert” and “20 years” are flattery aimed at something that can’t be flattered.
What does help is when a role carries genuine information about perspective:
✅ “Review this from the perspective of a sceptical CFO. What would they push back on?”
✅ “You’re a copy editor. Don’t rewrite — just flag what’s unclear.”
These work because they change what gets attended to. “Sceptical CFO” tells it to look for cost and risk. “Copy editor, don’t rewrite” sets a real boundary.
The test: does the role tell it something it didn’t already know? “Act as an expert writer” doesn’t — you already asked for good writing. “Act as someone who has to implement this on Monday” does.
Ask for the thinking
For anything involving judgement, comparison, or analysis, ask it to reason before concluding.
✅ “Before you recommend one, list the trade-offs of each option.”
✅ “Work through this step by step.”
Two reasons this helps. First, it genuinely improves the answer — the model builds on its own reasoning as it goes, so making the reasoning explicit gives it more to work with. Second, and more importantly for you: you can see where it went wrong. A bare conclusion is unauditable. A conclusion with reasoning attached lets you spot the bad assumption in step two.
Given Lesson 2 — that it’s confident whether or not it’s right — being able to inspect the reasoning is the closest thing you have to a check.
When to start over instead
Sometimes iterating makes things worse. You correct, it overcorrects, you correct back, and six messages later you’re further from what you wanted than when you started.
When that happens, start a fresh chat.
Remember from Lesson 2: the whole conversation gets re-sent every time. So every wrong turn is still sitting there, still being read, still influencing the output. You can’t take it back by saying “ignore that” — it’s still in the thread. The bad draft is now part of what it’s predicting from.
The signal to restart: you’re correcting the same thing twice. If you’ve said “less formal” two messages apart, the thread is poisoned. Open a new one, and this time put what you learnt into the first prompt.
That’s the real value of a failed conversation — it taught you what your prompt was missing.
Break big tasks into small ones
A prompt asking for a full report — research, structure, draft, polish — gets you a mediocre version of all four.
Split it:
- “What should a report on X cover for audience Y?”
- “Good. Turn that into an outline with section headings.”
- “Write section 2 only. 200 words.”
Slower. Far better. And you catch problems at step one, when fixing them costs nothing, instead of at the end when you’re staring at 2,000 words built on the wrong premise.
This maps onto how you’d brief a person, which is the useful intuition: nobody hands a new starter “write the report” and walks away.
Save what works
When you land on a prompt that produces exactly what you needed — keep it.
Not in your head. In a note, a doc, wherever. Most work is repetitive: the same weekly summary, the same kind of client email, the same tidy-up of the same messy input. You’ll want that prompt again next Tuesday.
This is the compounding bit. Six months in, the difference between someone who’s good at this and someone who isn’t isn’t talent. It’s that one of them has a folder of prompts that already work, and the other retypes from scratch every time and wonders why it’s slow.
Key takeaways
- Treat the first answer as a draft — expect roughly 60% and iterate from there
- Vague feedback gets vague changes; name exactly what’s wrong and keep what’s working
- “Act as a world-class expert” is cargo cult; roles only help when they add real perspective
- Ask for reasoning before conclusions — better answers, and you can see where it went wrong
- If you’re correcting the same thing twice, start a fresh chat instead
- Break big tasks into steps; catch problems early when they’re cheap
- Save prompts that work — this is where the compounding happens
Next: Lesson 6 – Using AI for Data and Analysis
You can prompt and iterate. Next: pointing it at information rather than a blank page — and where that gets dangerous.