You type: "Write an apology email for a late order." The reply comes back fine. It's also generic enough to paste into any support inbox on earth: no order number, no customer name, no real reason the delay happened. You tweak a sentence, send it, and wonder why the AI didn't just get it right the first time.
It didn't get it right because you didn't tell it enough, and there's a small set of ways to fix that. Not a hundred hacks, four techniques, each suited to a different kind of task. Here's how to pick.
For example: order #4471, a birthday gift, shipped 9 days late. The customer, Priya, emailed asking where it is. You need AI's help writing the reply and deciding what to offer her.
Give clear, specific instructions (Clear Instructions)
The vague version: "Write an apology email for a late order." The model has to guess the tone, the length, the customer's name, and what you're willing to offer. It guesses safely, which means generically.
The specific version:
Write a 120-word email to Priya apologizing that order #4471
(a birthday gift) shipped 9 days late. Offer a 20% discount
code for her next order. Sign it from the Support team at
Fernbrook Goods. Tone: warm, not corporate.
Same task, radically different output, because you removed the guessing. Clear Instructions means stating the constraints that matter: length, audience, specific facts, what "good" looks like. Reach for it first, every time, before anything fancier.
Show it examples (Few-Shot Prompting)
Clear instructions describe what you want. Sometimes tone is easier to show than describe. If you've written two replies you were happy with, paste them in and ask the model to match the pattern:
Here are two replies I've sent before that landed well:
1. [reply to a late-order complaint]
2. [reply to a damaged-item complaint]
Write a reply to Priya in the same voice: short sentences,
one specific detail about her order, no corporate filler.
This is Few-Shot Prompting: giving the model examples instead of, or alongside, a description. "Warm but not corporate" means something different to every writer; two real examples pin it down exactly.
Ask it to think step by step (Chain-of-Thought)
Writing the email is the easy part. Deciding whether Priya gets a discount or a full refund is a judgment call, and that's where a one-shot answer can quietly skip a step. Instead of asking for a verdict, ask for the reasoning:
Priya's order #4471 shipped 9 days late and missed a birthday.
Think through this step by step: what do we know about the
order and the customer, what would a refund cost us versus a
20% discount, and what would you recommend? Show your reasoning
before the recommendation.
Chain-of-Thought prompting asks the model to lay out its intermediate steps before landing on an answer. You often get the same recommendation either way, but now you can see if it weighed the right things. Use this for anything with more than one factor: pricing, triage, debugging a failure with several plausible causes.
Assign it a role (Role Prompting)
The email is written, the discount is decided. But should this reply sound like a founder personally handling a complaint, or a support rep following a script? That's a framing problem, and Role Prompting solves it by telling the model who it is before telling it what to do:
You are a senior customer support specialist at a small
e-commerce company. You're empathetic but concise, and you
always acknowledge the specific problem before offering
compensation.
A role doesn't add facts. It sets the lens the model reasons behind everything else, the cheapest way to keep a consistent voice across many prompts without restating the same style notes each time.
Which technique do you reach for?
These stack: a real prompt often uses two or three at once. But if you're starting from a vague prompt and picking one fix, match the task to the technique:
Your task | Reach for | Why |
|---|---|---|
A one-off output: an email, a summary, a short report | Clear Instructions | The model can't hit constraints you never gave it |
Matching a specific tone or format across many outputs | Few-Shot Prompting | Examples teach style faster and more precisely than description |
A decision with more than one factor: pricing, triage, debugging | Chain-of-Thought | Forces the reasoning into the open, so you can catch a bad step before it becomes a bad answer |
Consistent voice or judgment across a whole session | Role Prompting | Sets the lens once instead of restating it in every prompt |
If you're not sure, start with Clear Instructions. It fixes most mediocre answers on its own; the other three refine it, they don't replace it.
