How to Write Effective Prompts

1. Role Assignment

Assigning a role to the model shapes the vocabulary, depth, and framing of its response. A model told it is a senior security engineer will apply different heuristics than one given no context.

Before

Explain memory safety in C++.

After

You are a senior systems programmer with 15 years of C++ experience. Explain memory safety to a mid-level developer who is transitioning from Python. Cover RAII, smart pointers, and undefined behavior. Use short code examples.

The role does three things: sets expertise level, calibrates vocabulary to the audience, and implicitly constrains scope to what a person in that role would actually say.

2. Task Specification

Vague instructions produce vague results. The model will pick the most likely interpretation, which may not be yours. State your goal, constraints, and non-goals explicitly.

Summarize this document.

Summarize the following 5000-word product spec in 3 bullet points. Each bullet must be under 20 words. Focus on what changes, not what stays the same. Ignore the appendix.

Key constraints to specify:

• Length (word count, line count, number of items)
• Scope (what to include, what to exclude)
• Audience (developer, executive, end user)
• Tone (formal, concise, conversational)
• Format (bullet list, table, prose, JSON)

3. Few-Shot Examples

Few-shot prompting provides input/output examples before the real input. It is the most reliable way to communicate format and style without writing an exhaustive specification.

Classify the sentiment of customer feedback as POSITIVE, NEGATIVE, or NEUTRAL.

Input: "Shipping was fast and the product works great."
Output: POSITIVE

Input: "The packaging was damaged but the item inside was fine."
Output: NEUTRAL

Input: "Completely broken on arrival, terrible experience."
Output: NEGATIVE

Input: "{{feedback}}"
Output:

Three to five examples usually covers the space. More examples help with edge cases but increase cost. One example is often enough for format-only guidance.

4. Chain of Thought

For tasks that require reasoning—math, logic puzzles, multi-step decisions—asking the model to reason step by step before giving a final answer significantly improves accuracy. This is called chain-of-thought (CoT) prompting.

Is this Python function O(n²)?

def find_pairs(arr):
  result = []
  for i in range(len(arr)):
    for j in range(i+1, len(arr)):
      if arr[i] + arr[j] == 0:
        result.append((arr[i], arr[j]))
  return result

Is this Python function O(n²)?
Think step by step:
1. Identify each loop and its range
2. Determine how the ranges relate to n
3. Multiply loop complexities
4. State the final complexity with justification

def find_pairs(arr):
  result = []
  for i in range(len(arr)):
    for j in range(i+1, len(arr)):
      if arr[i] + arr[j] == 0:
        result.append((arr[i], arr[j]))
  return result

Common CoT triggers: “think step by step”, “reason through this”, “explain your reasoning before answering”, “work through each part”.

5. Output Formatting

Specifying the exact output format reduces post-processing and makes prompts suitable for programmatic use. The model can produce JSON, Markdown, CSV, and most other text formats reliably.

JSON output

Extract the key fields from the job posting below. Return a JSON object with these exact keys:
{
  "title": string,
  "company": string,
  "location": string,
  "remote": boolean,
  "salary_min": number | null,
  "salary_max": number | null,
  "required_skills": string[],
  "years_experience": number | null
}

Return only the JSON object. No explanation.

Job posting:
{{posting}}

Structured Markdown

Write a technical incident report with these sections:
## Summary
One sentence description of what happened.

## Timeline
Bullet list with timestamps.

## Root Cause
What caused the incident.

## Impact
Who was affected and for how long.

## Remediation
Steps taken to resolve.

## Prevention
Changes to prevent recurrence.

Incident details:
{{details}}

6. Common Mistakes

Don't use bullet points.

Use numbered paragraphs, one sentence each.

Summarize this article, extract action items, identify the author's bias, and suggest three follow-up questions.

Split into four separate prompts, or structure them as explicit numbered tasks with separate output sections.

Fix the bug in the same style as the rest of the codebase.

Fix the bug. Follow this style: 2-space indent, single quotes, no semicolons, arrow functions preferred.

List relevant papers on this topic.

List exactly 5 papers on this topic, ordered by relevance. Stop at 5.

7. Ready-to-Use Templates

The PromptIndex library contains tested templates for the most common use cases. Browse by category or start with these: