By the end of this lesson you will be able to
- Turn a vague request into a specific one using four fixed ingredients
- Get usable output in one or two attempts instead of six
- Know when to iterate and when to start over
Most disappointing AI output traces back to a request that would have confused a competent new hire too. "Write something about our new service" is not a brief. The fix is not a magic phrase — it is saying the things you were leaving implicit.
The four ingredients
Nearly every good prompt contains the same four things. Miss one and the model fills the gap with an average guess, which is exactly why the output feels generic.
Context — what is going on
Who you are, what the situation is, what has already happened. "We are a five-person plumbing company in Camrose. A customer is unhappy about a callout fee they say was not explained."
Task — what you want produced
One clear deliverable. "Draft a reply that keeps the fee but acknowledges we should have flagged it up front." Not "help me with this."
Format — what it should look like
Length, structure, medium. "Short email, four sentences, no bullet points, signed off as Dale." Format is the most commonly omitted ingredient and the one that saves the most rework.
Constraints — what to avoid
Tone limits, forbidden claims, things that must appear. "Do not apologise for the fee itself. Do not offer a discount. Mention the two-year warranty."
What that looks like in practice
The vague version
- "Write an email to an unhappy customer"
- Output is generic and over-apologetic
- Wrong length, wrong tone
- Four rounds of "no, more like…"
- Faster to have written it yourself
The specific version
- Context, task, format, and constraints stated
- Output is close on the first attempt
- Right length because you said the length
- One round of light editing
- Genuinely faster than writing it yourself
We're a five-person plumbing company in Camrose, Alberta.
A customer is unhappy about a $95 after-hours callout fee. The fee is
in our terms, but our tech didn't mention it on the phone.
Draft a reply from me (Dale, the owner).
Format: email, four to five sentences, plain language, no bullet points.
Constraints:
- Acknowledge we should have flagged the fee on the call
- Do not remove or discount the fee
- Do not be grovelling — friendly and straightforward
- Close by confirming the work is under our 2-year warrantyShow, do not describe
Describing a tone is unreliable — "professional but warm" means something different to everyone, including the model. Pasting an example of the thing you want is dramatically more effective than any adjective.
Iterating vs. starting over
When the first attempt misses, people tend to add a correction, then another, then another. Past about three corrections the conversation is usually carrying more confusion than signal — earlier wrong attempts are still in the context, still influencing the output.
A simple rule:
- Wrong details, right shape — correct it in place. That is what iteration is for.
- Right details, wrong shape — restate the format and ask for a rewrite.
- Wrong shape and wrong details after two attempts — start a new conversation with a better brief. You learned what was missing; use it.
Turning a good prompt into an asset
A prompt that worked is worth keeping. Most businesses do the same handful of writing tasks over and over — quotes, follow-ups, apologies, job descriptions, summaries. Five saved prompts covering your five most common tasks is worth more than any amount of prompting theory.
Do I need to be polite to it?
It makes no measurable difference to quality. Write however is natural to you — most people find that phrasing it as a real brief to a real person produces better briefs, which is the actual benefit.
Are those long prompt templates worth using?
Sometimes, for repeated high-value tasks. For everyday work the four ingredients cover almost all of the value. Treat elaborate templates as a sign a task has become routine enough to standardise, not as a starting point.
Why does the same prompt give different results?
The model samples from a distribution rather than looking up an answer — see the sampling demo in AI Foundations. If you need consistency, tighten the format constraints until variation has nowhere left to go.
Key takeaways
- Context, task, format, constraints. Missing one is why output feels generic.
- Format is the most-skipped ingredient and the biggest time-saver.
- Pasting an example beats describing a tone, every time.
- After two failed attempts, start over with a better brief instead of patching.
- Save the prompts that worked. Your five most common tasks are worth standardising.