The most common prompt engineering mistake is not being lazy. It is being vague. Vague prompts force the model to guess at what you mean. The guesses pile up and the output looks wrong.
These five prompt engineering best practices close most of that gap. Each one is concrete, takes under a minute to apply, and includes a before-and-after so you can see the difference.
Best practice 1: give the model measurable success criteria
This is the single biggest gap in most prompts. "Make it sound professional" gives the model nothing to optimize toward.
Before:
Make it sound professional.
After:
Use formal business language. No slang. Include technical terminology where appropriate. Active voice. End with two concrete next steps.
The rewrite is specific enough to argue with, which means it is specific enough to measure. If the model ignores one of those rules, you can point to it.
Best practice 2: assign a role and name the audience
The model has no idea who you are or who the output is for until you say so. Without that context, it picks the statistical average reader.
Before:
Explain quantum computing.
After:
You are a physicist explaining quantum computing to undergraduates who know basic linear algebra. Use one analogy from classical computing that lands for that audience. No more than 400 words.
The role does most of the work. Audience, depth, tone, and length all start to fall into place once the model knows who is in the room.
Best practice 3: specify output format
If you leave the format open, you get whatever the model felt like. Prose one time, bullets the next, JSON only if you are lucky.
Before:
Give me a response about AI.
After:
Return JSON with these fields:
- definition: max 100 words
- key_concepts: array of 5 terms with definitions
- real_world_apps: array of 3 examples with brief descriptions
- glossary: object mapping terms to definitions
For anything feeding a downstream script or pipeline, format is not optional. Specifying it once is far cheaper than reformatting by hand every run.
Best practice 4: include one example (few-shot)
This is the highest-leverage move on most struggling prompts. A single example shows the model register, tone, length, and structure without you having to describe any of them.
Before:
Translate these support tickets to Spanish.
After:
Translate these support tickets to Spanish.
Example: Input "My account is locked, please help." Output "Mi cuenta está bloqueada, por favor ayuda."
Now translate the following: ...
The example is doing more work than it looks like. It tells the model how literal to be, whether to preserve "please" as "por favor," and how to break sentences. None of that has to be spelled out as a rule.
Best practice 5: constrain the scope
A prompt that asks for too much produces a shallow response to everything. Pick fewer things and cover them well.
Before:
Write a blog post that explains AI to beginners, covering transformers, CNNs, RNNs, attention, gradient descent, backpropagation, loss functions, learning rates, overfitting, regularization, dropout, fine-tuning, quantization, and ethical considerations.
The model produces a textbook. Shallow on every topic.
After:
Explain AI to complete beginners in 600 words. Cover only:
1. What AI is (definition)
2. How it learns (training vs inference)
3. Two real-world applications
4. One common misconception to avoid
Use simple analogies. No jargon beyond "model" and "data."
A 600-word post on four ideas beats a 2,000-word post on twenty-four every time.
A prompt engineering checklist
Run any prompt through these five checks before you send it:
- Is the goal measurable? (not just "make it good")
- Have I named the role and the audience?
- Have I specified the output format?
- Have I included at least one example of what good looks like?
- Is the scope narrow enough to cover well?
Prompts that pass all five usually get a usable answer on the first try. Prompts that fail three or more almost never do.
A starter template
If you want a single starting structure:
ROLE: who the model should be
TASK: the one thing you want
AUDIENCE: who the output is for
RULES (must follow):
- hard constraint 1
- hard constraint 2
FORMAT: exact output shape
EXAMPLE: one example of what good looks like
Fill this in before sending and you will skip most of the "that's not what I wanted" rewrites.
A faster way to find the gaps
Paste any prompt into FixMyPrompt. The rubric scores all five areas above and tells you which one is costing you the most signal. The rewrite fills in the weak spots.
Three free reports per day. No signup.
Related reading
- The Prompt Grader Rubric: How We Score AI Prompts on a 0-100 Scale
- Why 'Rewriting' Your ChatGPT Prompt Doesn't Fix It
- Why ChatGPT Keeps Misunderstanding Your Instructions
- How to Make AI Writing Sound Less Like AI