Prompt Engineering as a Discipline

You asked an AI tool for something last week and got a great answer. Today you asked for something similar and got mush. Same tool, same you, very different result. It is tempting to decide the AI is moody, or that good results come down to luck.

They don't. Most of the difference between a useful answer and a useless one comes from what you gave the model to work with, and that is something you control. This lesson treats prompting like any other practical skill: a method, a way to check your work, and a habit of keeping what works.

What "discipline" actually means here

Prompt engineering sounds technical. In practice it is closer to writing clear instructions for a capable new coworker who knows a lot about the world and nothing about your situation. They will do exactly what you describe, fill every gap with a guess, and never ask a follow-up question unless you invite one.

A discipline is just a repeatable way of working. For prompting, it comes down to four habits:

1. Test. Run the prompt and look at what you actually got, not what you hoped for.
2. Compare. Change one thing, run it again, and put the two results side by side.
3. Iterate. Keep the change if it helped, undo it if it didn't, and try the next idea.
4. Save what works. When a prompt reliably does its job, store it so you never have to rebuild it from memory.

None of this requires code or special software, just treating a prompt as a draft instead of a wish.

Why results feel random

A few things make AI output feel unpredictable even when nothing mysterious is happening.

The model fills gaps with averages. If you ask for "a cover letter," you get the most typical cover letter the model can produce, because you gave it nothing that points anywhere else. Generic in, generic out.

The same prompt can give different answers. Most AI tools add some variation on purpose, so running an identical prompt twice rarely gives identical wording. That is one reason a single good result proves very little. You need to see a prompt work more than once before you trust it.

Conversation history leaks in. Earlier messages in the same chat shape later answers. A prompt that worked in a long conversation may behave differently in a fresh one, because the earlier context was quietly doing some of the work.

You changed more than you think. "I asked the same thing" often means "I asked roughly the same thing." One missing detail, like who the audience is, can change the whole answer.

Once you know these causes, you can control for them: be specific, test more than once, start fresh chats when you want a clean test, and change one thing at a time.

Start with what you needed, not what you got

The most important step happens before you write the prompt. Decide what a good answer looks like. If you skip this, you will judge the output by whether it sounds impressive, and AI output almost always sounds impressive.

Write down two or three checks, in plain words. For example, if you need a reply to an upset customer:

• It apologizes once, without groveling.
• It states the exact next step and when it will happen.
• It is short enough to read on a phone.

Now you have something to measure against. When the output arrives, you are not asking "is this good?" You are asking "does it pass my three checks?" That question has an answer.

A simple way to evaluate an output

Use these five questions on anything the AI gives you. You don't need all of them every time, but running through them takes a minute and catches most problems.

The last two rows matter most. AI tools can state wrong information in a confident tone, and they will sometimes invent a detail, a quote or a source to make an answer feel complete. You are the checker. Anything important, such as a figure going into a report, a medical or legal point, or a claim with your name on it, gets verified by a person before it is used.

Worked example: one prompt, three rounds

Say you run a small tutoring service and want a short text message reminding parents about a schedule change.

Round 1.

Write a reminder about the schedule change.

The result is a formal paragraph that mentions "the upcoming schedule change" without saying what it is, because you never said. It fails the first check: it doesn't do the job.

Round 2. Add the facts.

Write a text message to parents. Tutoring on [DAY] is moving from
[OLD TIME] to [NEW TIME], starting [DATE]. The location stays the same.

Better. The facts are there. But it is four sentences long with a cheerful sign-off, and parents will read it on a phone between other things. It fails the shape check.

Round 3. Change one thing: the constraint on length and tone.

Write a text message to parents. Tutoring on [DAY] is moving from
[OLD TIME] to [NEW TIME], starting [DATE]. The location stays the same.
Keep it under 40 words, friendly but plain, and end by asking them
to reply "OK" to confirm.

Now it passes. Notice what happened: each round changed one thing, and you could see exactly which change fixed which problem. If you had rewritten everything at once, you would not know what worked.

Run the final version two or three more times to make sure it holds up. If it does, save it. The next time the schedule changes, you fill in the brackets and you're done in seconds.

Save what works

A saved prompt is worth more than a good memory of one. Keep a simple document or note with:

• A name that says what it does, such as "Parent schedule text."
• The prompt itself, with [BRACKETED] spots for the parts that change.
• One line on when to use it and anything it tends to get wrong.

This is the start of a personal prompt library, and it grows every time a prompt earns its place.

Try this

Set aside 10 to 15 minutes.

1. Pick a real task you did recently with AI, or one you have coming up this week.
2. Before prompting, write two or three checks that a good answer must pass.
3. Write your first prompt the way you normally would. Run it and score it against your checks.
4. Change one thing only: add a missing fact, name the audience, or set a length. Run it again in a fresh chat and compare.
5. Repeat once more. If you have access to the Study AI Mastery Playground, try running your final prompt on two or three models side by side in compare mode and note where they differ.
6. Save the version that passed, with brackets for the parts that change.

What to remember

• Results that feel random usually trace back to gaps in the prompt, and gaps are something you can fix.
• Decide what a good answer looks like before you prompt, so you judge the output against your needs, not against how polished it sounds.
• Change one thing at a time so you know which change made the difference.
• Test a prompt more than once before you trust it, because the same prompt can produce different answers.
• AI output can be wrong or invented. A person checks anything important before it is used.
• Save prompts that work, with placeholders, so good results become repeatable.