Prompting Basics
Fundamental Prompting Principles
This guide is based on Meta’s official documentation and best practices from the prompting community. For more details, see the Llama prompting guide.
Prompt engineering is the practice of crafting effective inputs to get the best possible outputs from language models. It involves providing context, instructions, and examples to guide the model toward your desired result.
The Anatomy of a Good Prompt
A well-crafted prompt typically includes:
- Context: Background information the model needs
- Instruction: What you want the model to do
- Input Data (optional): Text or information to process
- Output Format (optional): How you want the response structured
- Constraints (optional): Limitations or requirements
Example Structure:
[Context] You are an expert librarian helping students.
[Instruction] Extract all books from the following text.
[Input Data] {text about books}
[Output Format] Return as a bullet list with: Title, Author, Year.
[Constraint] Only include books published after 1900.Core Prompting Principles
1. Be Clear and Specific
Poor Prompt:
Tell me about climate change.Better Prompt:
Explain the primary causes of climate change in simple terms suitable
for a high school student. Focus on the top 3 factors and their impacts.
Limit your response to 200 words.Why It’s Better:
- Specifies audience level (high school)
- Defines scope (top 3 factors)
- Sets length constraint (200 words)
- Clarifies purpose (explain causes and impacts)
2. Provide Context
The model doesn’t know who you are, what you’re working on, or what you already know. Provide necessary context.
Without Context:
How do I organize this?With Context:
I'm a librarian organizing a collection of 500 historical photographs
from the 1920s-1940s. How should I categorize and catalog these images
to make them easily searchable for researchers?Why It’s Better:
- Model knows your role (librarian)
- Understands the task (organizing photos)
- Has details (time period, quantity, purpose)
3. Use Examples (Few-Shot Learning)
Show the model what you want by providing examples.
Zero-Shot (no examples):
Extract book information from this text.Few-Shot (with examples):
Extract book information from text and format as shown:
Example 1:
Input: "George Orwell's 1984 was published in 1949."
Output: Title: 1984 | Author: George Orwell | Year: 1949
Example 2:
Input: "To Kill a Mockingbird by Harper Lee came out in 1960."
Output: Title: To Kill a Mockingbird | Author: Harper Lee | Year: 1960
Now extract from this text:
"John Steinbeck wrote Of Mice and Men, published in 1937."Expected Output:
Title: Of Mice and Men | Author: John Steinbeck | Year: 19374. Specify Output Format
Tell the model exactly how you want the response structured.
Unspecified Format:
List the books mentioned in this text.Specified Format:
List the books mentioned in this text in the following format:
```json
{
"books": [
{
"title": "Book Title",
"author": "Author Name",
"year": 1234,
"language": "English"
}
]
}Return only valid JSON, no additional text.
---
### 5. Set Constraints and Boundaries
Define what the model should and shouldn't do.
**Example with Constraints**:Summarize this research paper.
Requirements:
- Maximum 150 words
- Focus only on methodology and findings
- Do not include author opinions or speculation
- Write in present tense
- Use bullet points for key findings
---
## Practical Examples with Sample Text
Let's use the following sample text for our examples:John Steinbeck’s “Of Mice and Men” was published in English in 1937. Gabriel García Márquez wrote Cien años de soledad, a Spanish-language novel, published in 1967. “Pride and Prejudice” by Jane Austen came out in 1813 in English. Haruki Murakami penned the Japanese novel ノルウェイの森 (Norwegian Wood), published in 1987.
### Example 1: Simple Extraction
**Prompt**:Extract all book titles from the following text:
{insert sample text above}
List only the titles, one per line.
**Expected Output**:Of Mice and Men Cien años de soledad Pride and Prejudice ノルウェイの森 (Norwegian Wood)
---
### Example 2: Structured Data Extraction
**Prompt**:Extract book information from the following text and return as a table:
{insert sample text}
Create a markdown table with columns: Title | Author | Year | Language
**Expected Output**:
```markdown
| Title | Author | Year | Language |
|-------|--------|------|----------|
| Of Mice and Men | John Steinbeck | 1937 | English |
| Cien años de soledad | Gabriel García Márquez | 1967 | Spanish |
| Pride and Prejudice | Jane Austen | 1813 | English |
| ノルウェイの森 (Norwegian Wood) | Haruki Murakami | 1987 | Japanese |Example 3: Filtering with Constraints
Prompt:
From the following text, list only books published after 1900:
{insert sample text}
Format: "Title" by Author (Year)Expected Output:
"Of Mice and Men" by John Steinbeck (1937)
"Cien años de soledad" by Gabriel García Márquez (1967)
"ノルウェイの森" by Haruki Murakami (1987)Example 4: Creative Rewriting
Prompt:
Rewrite the following text as if it's being explained to a 10-year-old
child who loves reading:
{insert sample text}
Use simple language, short sentences, and an enthusiastic tone.Expected Output (example):
Wow, there are so many cool books to talk about! Let me tell you about them!
First, there's a book called "Of Mice and Men" written by a man named
John Steinbeck. It came out way back in 1937. That's a really long time ago!
Then there's this amazing Spanish book called "Cien años de soledad"
(that means "One Hundred Years of Solitude" in English!). A famous
writer named Gabriel García Márquez wrote it in 1967.
Have you ever heard of "Pride and Prejudice"? It's a super old book
from 1813, written by Jane Austen. It's in English and lots of people
still love reading it today!
And there's a cool Japanese book called "Norwegian Wood" (which is
"ノルウェイの森" in Japanese). A writer named Haruki Murakami wrote
it in 1987.Stylization Techniques
Academic Style
Prompt:
Analyze the following text from an academic perspective, focusing on
the geographic and temporal distribution of literary works:
{insert sample text}
Use formal academic language and cite specific examples.Technical/Structured Output
Prompt:
Convert the following text into a JSON array of book objects:
{insert sample text}
Each object should have: title, author, year, language, and originalTitle
(if different from English title).
Return only valid JSON.Creative/Storytelling Style
Prompt:
Using the books mentioned in this text, write a short story about a
magical library where each book represents a different world:
{insert sample text}
Maximum 200 words. Make it whimsical and engaging.Common Prompting Pitfalls
❌ Pitfall 1: Being Too Vague
Bad:
summarize thisGood:
Summarize the following text in 3 bullet points, focusing on the
main themes:❌ Pitfall 2: Assuming Context
Bad:
How do I fix this?Good:
I'm seeing an error "ModuleNotFoundError: ollama" when running my
Python script. I'm using Python 3.11 on macOS. How do I fix this?❌ Pitfall 3: No Output Format
Bad:
Give me the dataGood:
Extract the data and format as CSV with headers: title, author, year❌ Pitfall 4: Multiple Unrelated Tasks
Bad:
Summarize this, translate it to French, and write a poem about it.Good: Break into separate prompts:
1. First: Summarize this text in 2-3 sentences.
2. Then: Translate the summary to French.
3. Finally: Write a haiku inspired by the main theme.Best Practices Checklist
Before submitting your prompt, ask:
- Is my instruction clear and specific?
- Have I provided necessary context?
- Have I specified the output format?
- Have I set appropriate constraints?
- Am I asking for one task or multiple related tasks?
- Have I provided examples if the task is complex?
- Is the prompt self-contained (doesn’t assume prior knowledge)?
Iterative Refinement
Prompting is an iterative process. Here’s how to refine:
- Start Simple: Begin with a basic prompt
- Evaluate Output: Check if it meets your needs
- Identify Gaps: What’s missing or wrong?
- Refine: Add specificity, examples, or constraints
- Test Again: Try the improved prompt
- Repeat: Keep refining until satisfied
Example Iteration:
Version 1 (too vague):
List the books.Version 2 (better, but no format):
List all books mentioned in this text.Version 3 (specifies format):
List all books mentioned in this text in the format: Title by Author (Year).Version 4 (adds constraints and sorting):
List all books mentioned in this text in the format: Title by Author (Year).
Sort chronologically by year, oldest first.
Only include books published before 1950.Next Steps
Now that you understand the basics, let’s explore advanced techniques: