Two prompts, one question
Someone asks:
How do I improve my CV?
They get a list. Use action verbs. Quantify your achievements. Tailor it to the role. Keep it to one page. All true, all useless, because they already knew it.
Now the same person asks:
I am applying for a junior data analyst job at a bank in Lagos. My CV is pasted below. The advert asks for SQL, Excel and "stakeholder communication". I have no paid experience — only a six-month university project where I cleaned and analysed 40,000 hospital records in Postgres. Rewrite the top third of the CV so a recruiter sees the SQL work in the first ten seconds. Keep the whole thing to one page. Do not invent any experience I have not described.
[full CV pasted]
Both people want a better CV. The answers are not remotely comparable, because only the second question is answerable.
Why that happens
A language model produces a plausible continuation of the text in front of it. That is genuinely most of what is going on. It has no picture of your life, your file, or your Thursday deadline. It has your words and nothing else.
"How do I improve my CV?" is a question that thousands of articles have already answered. The most plausible continuation is a generic list, because a generic list is what generically follows that sentence. The model is not being lazy. It is being accurate about a vague question.
The second prompt cannot be continued generically. There is one CV, one advert, one awkward gap to cover.
Average in, average out. When someone says a model gave them a bland answer, they have usually asked a question whose honest answer is bland.
What actually did the work
Look at what the second prompt added:
- Who and where. Junior, Lagos, banking. What a CV should look like changes by market.
- The target. The advert's own three words, quoted.
- The raw material. The CV itself, pasted, not described.
- The hard part. No paid experience. This is the difficulty of the task. Hiding it would have produced advice for someone else.
- A finished state. Top third rewritten, one page, SQL visible in ten seconds.
- A boundary. Do not invent experience.
None of that is a trick. It is the briefing you would give a friend who offered to help over lunch.
The incantations are mostly not the thing
A lot of prompting advice is about magic words. Tell it that it is a world-class expert. Offer it a tip. Tell it to take a deep breath. Threaten it politely.
Some of those had measurable effects on the weaker models of 2023. Most have shrunk or disappeared as models improved, and several were never more than folklore repeated between blog posts. You will still find them recommended confidently today.
The honest version: the reliable levers are what you are trying to do, the material you are working from, the constraints that make it hard, and the shape you want back. Everything else is decoration.
That is good news. There is no secret language to learn. There is a habit of under-describing your own problem, and you can drop it.
A test that takes one minute
Before you send a prompt, read it as if a competent stranger had sent it to you, with no memory of your week. Ask two questions:
- Could I start this task from this text alone?
- What would I have to ask you first?
Every question you would have to ask is a missing line. Add those lines and send it again.
That is the whole method. The rest of this course is about doing each part of it well.
Before you move on