tools-only Prompt Engineering
A systematic approach to writing effective prompts for AI systems.
Core Principles
1. Be Specific, Not Vague
❌ Bad: "Write something about dogs" ✅ Good: "Write a 200-word blog post about the benefits of adopting senior dogs, targeting first-time dog owners"
2. Provide Context
❌ Bad: "Fix this code" ✅ Good: "Fix this TypeScript function that calculates user subscription costs. It should handle monthly and annual billing, with a 20% discount for annual plans. Current bug: returns NaN for annual subscriptions."
3. Specify Format
❌ Bad: "Give me a summary" ✅ Good: "Summarize this article in 3 bullet points, each under 20 words"
Prompt Structure
Basic Template
[ROLE/CONTEXT]
You are a [role] helping with [task].
[TASK]
Your task is to [specific action].
[INPUT]
Here is the [input type]:
{input}
[FORMAT]
Respond in the following format:
- [format specification]
[CONSTRAINTS]
Important rules:
- [constraint 1]
- [constraint 2]
Example
You are a senior software engineer reviewing code for security vulnerabilities.
Your task is to identify potential security issues in the following code and suggest fixes.
Here is the code to review:
{code}
Respond in the following format:
1. **Issue**: [description]
**Severity**: [Critical/High/Medium/Low]
**Fix**: [suggested fix]
Important rules:
- Focus only on security issues, not style
- Prioritize by severity
- Include code examples for fixes
Techniques
Chain of Thought
For complex reasoning, ask for step-by-step thinking:
Solve this problem step by step:
1. First, identify what we know
2. Then, determine what we need to find
3. Apply the relevant formula/logic
4. Calculate the answer
5. Verify the result
Few-Shot Examples
Provide examples of desired output:
Convert these sentences to formal business language.
Example 1:
Input: "Hey, can we chat about the project later?"
Output: "I would like to schedule a meeting to discuss the project at your earliest convenience."
Example 2:
Input: "This deadline is crazy tight!"
Output: "The timeline for this deliverable presents significant challenges."
Now convert:
Input: "{user_input}"
Negative Examples
Show what NOT to do:
Write a professional email response.
DO NOT:
- Use casual language like "Hey" or "What's up"
- Include emojis
- Be overly wordy
DO:
- Use a formal greeting
- Be concise and clear
- Include a clear call to action
Role Assignment
Assign a specific persona:
You are a Kubernetes expert with 10 years of experience in production deployments.
You explain concepts clearly to developers who are new to container orchestration.
You always provide practical examples alongside theory.
Prompt Patterns
Classification
Classify the following customer message into one of these categories:
- billing
- technical_support
- feature_request
- general_inquiry
Message: "{message}"
Respond with only the category name.
Extraction
Extract the following information from this job posting:
- Job title
- Required years of experience
- Required skills (as a list)
- Salary range (if mentioned)
Job posting:
{posting}
Return as JSON.
Generation with Constraints
Write a product description for {product}.
Constraints:
- Exactly 50-75 words
- Include 2 key benefits
- End with a call to action
- Use active voice
- Target audience: {audience}
Analysis
Analyze this code for performance issues.
For each issue found:
1. Describe the problem
2. Explain the impact
3. Suggest an optimized solution
4. Estimate the improvement
Code:
{code}
Testing Prompts
Test Cases
Always test with:
- Happy path — Normal, expected input
- Edge cases — Empty, very long, special characters
- Adversarial — Attempts to break or manipulate
- Domain-specific — Industry/context-specific scenarios
Iteration Loop
1. Write initial prompt
2. Test with 5+ varied inputs
3. Identify failure modes
4. Refine prompt
5. Repeat until consistent
Common Mistakes
| Mistake | Fix |
|---|---|
| Too vague | Add specific details and constraints |
| Too long | Remove unnecessary context, focus on essentials |
| No examples | Add 2-3 clear examples |
| Ambiguous output format | Specify exact format (JSON, bullets, etc.) |
| Missing edge cases | Add explicit handling for edge cases |
| Assuming knowledge | Provide necessary context |
System Prompt Best Practices
For AI agents/assistants:
You are [role] for [company/product].
## Core Behavior
- [Key behavior 1]
- [Key behavior 2]
## Capabilities
You CAN:
- [Capability 1]
- [Capability 2]
You CANNOT:
- [Limitation 1]
- [Limitation 2]
## Response Style
- [Style guideline 1]
- [Style guideline 2]
## Important Rules
- [Critical rule 1]
- [Critical rule 2]
Evaluation Criteria
Rate prompts on:
- Clarity — Is the task unambiguous?
- Completeness — Is all necessary context provided?
- Specificity — Are output requirements clear?
- Robustness — Does it handle edge cases?
- Efficiency — Is it as concise as possible?