Lesson 4 of 8 · 9 min read
Two people open the same AI tool. One gets something they can use in thirty seconds. The other gets bland, generic mush, decides AI is overhyped, and stops.
The difference is almost never the tool. It’s what they typed.
This is the lesson where the course stops being background and starts being a skill. And the skill is smaller than it sounds — there’s no secret syntax, no magic words, no list of “1,000 prompts that will 10x your productivity.” There are four things a good prompt contains. That’s the whole thing.
Why prompting is a skill at all
Remember Lesson 2: the model predicts what a good answer looks like based on what’s in front of it. That’s not a limitation to work around — it’s the mechanism you’re steering.
If you type “write an email about the project delay,” the model has almost nothing to go on. Which project? Delayed how long? Writing to your boss or a client? Are you apologising or informing? Is this a disaster or a minor slip?
It doesn’t know. So it predicts the average of every project-delay email it has ever seen. And the average of everything is, by definition, generic. That’s not the model failing. That’s the model doing exactly what you asked with exactly what you gave it.
Vague in, average out. Every time.
The four ingredients
Here’s what separates a useful answer from a useless one.
1. Context
Who you are, who it’s for, and what the situation is.
This is the ingredient people skip, and it’s the one that changes the output the most. The model has no idea about your job, your company, your reader, or your history with them. Everything obvious to you is invisible to it.
❌ “Write an email about the project delay.”
✅ “I’m a project manager at a software company. Our client’s website redesign is running two weeks late because we lost a developer. The client is important, generally reasonable, but has already been patient once. I need to tell them.”
Same request. The second one can actually be answered.
2. The task, specifically
What exactly do you want it to do?
“Help me with this” isn’t a task. “Improve this” isn’t a task. Be precise about the verb: summarise, rewrite, critique, list, compare, explain, draft.
❌ “Help me with this report.”
✅ “Summarise this report into five bullet points that a non-technical director could understand in one minute.”
Notice the second contains the purpose, not just the instruction. “For a non-technical director, in one minute” tells the model what to leave out — which is most of what a good summary is.
3. Format
What shape should the answer come in?
Left unsaid, you’ll get the default: a few paragraphs of readable prose, medium length, slightly over-explained. Fine sometimes. Often not what you wanted.
✅ “Give me a table with three columns: risk, likelihood, mitigation.”
✅ “Three bullet points, maximum ten words each.”
✅ “Write it as a Slack message, not an email. Casual, under fifty words.”
This is the cheapest win in the whole lesson. One sentence, dramatically better output.
4. Examples
Show it what good looks like.
This is the most powerful ingredient and the least used. If you have something written the way you want, paste it in.
✅ “Here are two product descriptions we’ve published before. Match their tone and length exactly. [paste] Now write one for our new item: [details]”
Describing a tone rarely works — “professional but friendly” means nothing in particular, and the model will produce the average of everything ever described that way. Showing a tone works immediately. One good example beats a paragraph of adjectives.
Before and after
Watch what happens when you stack all four.
Before:
“Write a job ad for a marketing person.”
You’ll get something. It’ll mention “dynamic team player” and “fast-paced environment” and read like every job ad on the internet, because that’s precisely what it’s the average of.
After:
“I run a 12-person B2B software company in Reading. I’m hiring our first marketing hire — someone who’ll do everything from content to events, reporting to me directly.
Write a job ad for it.
Around 300 words, structured as: what the role is, what you’ll do, what we’re looking for, what we offer. Plain English — no ‘dynamic team player’, no ‘fast-paced environment’. Honest about the fact that it’s a broad role at a small company.
For tone, match this ad we ran last year: [paste]”
Same tool, same model, thirty seconds more typing. Completely different result — and one you might actually publish.
That gap is the entire skill. There isn’t a level beyond it that requires special knowledge.
The mistakes almost everyone makes
Treating it like a search engine. You’ve spent twenty years learning to type the fewest possible words into Google. That instinct is precisely backwards here. Search rewards brevity. Prompting rewards context. Unlearning this is most of the battle.
Assuming it knows things it can’t know. Your company, your project, last week’s meeting, the document open in your other tab. If you didn’t paste it in, it doesn’t exist.
Being polite instead of clear. “Could you possibly help me maybe improve this a bit?” isn’t kinder — it’s just vaguer. Direct instructions aren’t rude to a model. Say what you want.
Asking for everything at once. A prompt containing six unrelated requests gets six mediocre answers. Do one thing well, then the next.
Accepting the first answer. This is the big one, and it gets Lesson 5 to itself.
A template worth stealing
When you’re stuck, this covers it:
Context: I’m [role] at [type of organisation]. [Relevant situation.]
Task: [Specific verb] [specific thing].
Audience: This is for [who], who needs [what].
Format: [Length, structure, tone.]
Example: Here’s something in the style I want: [paste]
You won’t need every field every time. A quick rewrite doesn’t need an example. But when an output disappoints, run down the list — the problem is almost always a missing ingredient, and you’ll usually spot which one in about five seconds.
That diagnosis is the real value here. Not the template. The ability to look at a bad answer and know why it was bad.
Key takeaways
- Bad output is usually a bad prompt, not a bad tool
- Vague in, average out — the model predicts from what you give it, so thin input gives you the average of everything
- Four ingredients: context, a specific task, a format, and ideally an example
- Showing beats describing — one pasted example outperforms any amount of adjectives
- Stop typing like it’s Google; search rewards brevity, prompting rewards context
- When an answer disappoints, work out which ingredient was missing
Next: Lesson 5 – Prompt Engineering That Works
You can now write a solid first prompt. Next: what to do when the first prompt isn’t enough — which is most of the time, and isn’t a failure.