Chrome Extension
ChatGPT only gives you answers as good as the prompt you type. Same model, different prompt, completely different result.
Once you know the different types of ChatGPT prompts, you stop guessing and start picking the right one for the job.
In this guide I'll walk you through 12 prompt types, when I use each one, and a copy-paste template for every type.
Whether you write content, run marketing, or write code, this will save you a lot of back-and-forth with the chatbot.
Key Takeaways
Understanding ChatGPT Prompts
What Are ChatGPT Prompts?
A ChatGPT prompt is simply the instruction you give a large language model (LLM). It can be one line or a full page, depending on what you need.
Your job is to give it enough context and direction so the answer is useful the first time.
The Importance of Context
Context is what turns a generic answer into an accurate one. When ChatGPT knows the background of your task, it guesses less and gets more precise.
Types of Prompts
Zero-Shot Prompting
Zero-shot prompting means you give ChatGPT a task with no examples at all. It relies on what the model already learned in training.
Task: {WHAT YOU WANT DONE}
Context: {WHO/WHAT/WHY, any key background}
Audience: {BEGINNER / EXEC / DEVELOPER / CUSTOMER}
Output format: {BULLETS / TABLE / JSON / STEPS / EMAIL / SCRIPT}
Constraints: {WORD COUNT, TONE, MUST/AVOID, TOOLS, SOURCES}
Success criteria: {WHAT A “GOOD” ANSWER LOOKS LIKE}Few-Shot Prompting
Few-shot prompting means you show ChatGPT a few examples first. It picks up the pattern and follows it, so the output is more accurate and on-target.
You are a performance copywriter.
Write {COPY TYPE} for {PRODUCT/OFFER}.
Constraints:
Hook style: {HOOK TYPE}
Benefit bullets: {COUNT}
CTA: {CTA}
Avoid: {BUZZWORDS/CLAIMS/ETC}
Examples (match the rhythm and punchiness)
Example 1
Input: {PRODUCT_1 + CONTEXT}
Output:
{COPY_1}
Example 2
Input: {PRODUCT_2 + CONTEXT}
Output:
{COPY_2}
Now write this
Input: {PRODUCT + CONTEXT}
Output:Chain-of-Thought Prompting
Chain-of-thought prompting walks the AI through a logical sequence of steps to reach an answer. I use it for anything that needs reasoning or problem-solving.
Do {TASK} and provide a clear, teachable explanation.
Requirements:
Give a concise answer.
Then give a step-by-step explanation that is easy to follow, including any intermediate results that matter.
Do not include hidden internal reasoning, only the reasoning needed for clarity.
Output:
Answer:
Explanation:Self-Consistency Prompting
Self-consistency prompting asks for several solutions to the same problem and then picks the answer that shows up most consistently.
Think of it as an upgrade to step-by-step reasoning when you need more reliable answers.
You are an expert {ROLE}.
Task: {PROBLEM}
Generate {N} independent solutions.
Rules:
- Use different approaches if possible.
- For each attempt, provide: (a) short reasoning summary, (b) final answer only.
Then:
- Compare the {N} final answers.
- Choose the most consistent answer.
- Output ONLY:
1) Final Answer
2) Confidence (0 to 100)
3) 1 or 2 line justification for why this answer won.
Least-to-Most Prompting
Least-to-most prompting breaks a hard problem into smaller, easier pieces. The AI solves them in order and uses earlier answers to help with later steps.
You are {ROLE}. Solve the task using least-to-most decomposition.
Problem: {PROBLEM}
Step 1 (Decompose):
- List the smallest set of subproblems needed to solve the main problem.
- Keep each subproblem simple and answerable.
Step 2 (Solve):
- Solve subproblem 1, then 2, etc.
- Reuse prior sub-answers.
Output format:
- Subproblems:
1) ...
2) ...
- Solutions:
1) ...
2) ...
- Final answer: ...
- Quick check: ...
Self-Ask Prompting
Self-ask prompting makes the model write the follow-up questions it needs, answer them, and then put together the final response.
It works really well with search or retrieval when facts matter.
You are an expert {ROLE}.
Main question: {QUESTION}
First, write "Follow-up Questions:" and list 3 to 7 questions you must answer to solve this well.
Then write "Answers:" and answer each question (use the provided context, or note what is missing).
Then write "Final:" and provide the final answer.
Constraints:
- If a follow-up requires external data and none is provided, say: "Missing data: {WHAT}" and continue.
- Keep the final answer clean and user-ready.
ReAct Prompting
ReAct mixes reasoning with actions (like using tools, search, or a knowledge base). The model thinks, does something, checks the result, and adjusts.
You are {ROLE}. Use Reason + Act steps.
Goal: {GOAL}
Available tools/actions: {TOOLS_LIST}
Loop:
1) Reason: Briefly state what you will do next and why.
2) Act: Choose ONE action from {TOOLS_LIST} and specify exact input.
3) Observe: Summarize what you learned from the result.
Repeat until solved.
Final:
- Answer:
- Evidence used:
- Remaining uncertainties (if any):
Tree-of-Thoughts Prompting
Tree-of-thoughts explores several possible reasoning paths, scores them, and picks the best route instead of following one straight line.
You are {ROLE}. Solve using a tree of options.
Problem: {PROBLEM}
1) Generate {K} candidate "thoughts" (approaches).
2) Score each thought from 1 to 10 using criteria: {CRITERIA}.
3) Expand the top {TOP_M} thoughts into next-step thoughts.
4) Repeat for {DEPTH} rounds or until a solution is clear.
5) Choose best path and produce final output.
Output format:
- Candidates + scores:
- Best path (high-level):
- Final answer:
- Why this path won:
Retrieval-Augmented Prompting (RAG-style Prompting)
RAG-style prompting gives the model your documents and tells it to answer using only that evidence. It's the standard way to get factual, reliable answers.
You are {ROLE}. Answer using ONLY the provided sources.
Question: {QUESTION}
Sources:
[1] {DOC_1}
[2] {DOC_2}
...
Rules:
- If the answer is not in the sources, say "Not found in sources."
- Quote or cite the source number for each key claim.
- Do not use outside knowledge.
Output:
- Answer:
- Source citations:
- Gaps / what to retrieve next:
Self-Refine Prompting
Self-refine is a loop: write a draft, critique it, then rewrite it using that critique. You can repeat it for a few rounds.
You are {ROLE}. Use iterative self-refinement.
Task: {TASK}
Constraints: {CONSTRAINTS}
Quality bar: {RUBRIC}
Round 1:
- Draft:
Round 1 Critique:
- Issues (bullet list):
- What to improve (bullet list):
Round 2:
- Revised Draft:
Stop after {ROUNDS} rounds or when rubric is satisfied.
Final output: provide only the best draft, plus a 3-bullet changelog.
Structured Output Prompting (Schema-First)
Structured output prompting forces the answer into a strict shape (usually JSON). That makes it reliable to plug into other tools and automations.
Return output in EXACT JSON matching this schema:
{PASTE_JSON_SCHEMA_OR_FIELD_LIST}
Task: {TASK}
Input: {INPUT}
Rules:
- Do not add extra keys.
- Use empty strings or null when unknown (as allowed).
- Ensure enums and types match.
Return JSON only.
Program-Aided Prompting (PAL-style)
PAL-style prompting asks the model to write code (like Python) for the parts it would likely get wrong in plain text, then use the computed result in its answer.
You are {ROLE}. Use program-aided reasoning.
Problem: {PROBLEM}
Data: {DATA}
1) Write Python code to compute the answer.
2) Show the computed result.
3) Explain the result in plain English.
4) Provide the final answer.
Output format:
- Code:
- Result:
- Explanation:
- Final:
Crafting Effective Prompts
Clarity and Specificity
Clarity and specificity matter more than anything else. A vague prompt gets you a vague, off-target answer.
Tone of Voice and Style
The tone you ask for shapes the tone you get. Set it on purpose.
Practical Tips and Examples
Actionable Tips for Prompt Engineering
Examples of Effective Prompts
A Before-and-After Prompt Example
Here's what all of this looks like on a real task. Same request, two prompts.
Before: the vague prompt
Write a LinkedIn post about time management.
This is zero-shot with no context. ChatGPT has to guess the reader, the length, and the tone, so you get something generic.
After: the clear prompt
Write a LinkedIn post about time management for early-career project managers in tech. Goal: encourage readers to try one practical method this week Context: they feel overwhelmed by meetings and Slack messages Tone: supportive, confident, not preachy Constraints: 120 to 160 words, no hashtags, no emojis, avoid buzzwords Output format: - 1-sentence hook - 3 bullet tips - 1 closing question
The second prompt adds the context, constraints, and output format we covered above. Now ChatGPT knows who it's writing for and what shape the post should take, so there's far less to fix afterwards.
The same trick works for job hunting too. If you're updating your CV, my ChatGPT prompts for resume writing are a good place to practice it.
A Reusable Prompt Template
When I don't know which type to start with, I fill in this template. It covers most everyday tasks, and you can bolt any of the 12 types on top of it.
Goal: [WHAT YOU WANT] because [WHY YOU NEED IT] Context: [BACKGROUND / DETAILS / SOURCE TEXT] Audience: [WHO IT'S FOR + KNOWLEDGE LEVEL] Constraints: [LENGTH] + [TONE] + [DO/DON'T] + [SCOPE] + [SOURCES IF NEEDED] Output format: [BULLETS / TABLE / OUTLINE / STEPS] Example to imitate: [PASTE A SHORT SAMPLE]
Add an example at the bottom and it becomes few-shot. Add "solve it step by step" and it becomes chain-of-thought. Simple as that.
Want to see this kind of structure filled in for real tasks? I put 100+ of them in my AI prompt examples guide.
Set It Once With ChatGPT Custom Instructions
If you keep typing the same context into every prompt (your job, your audience, the tone you like), stop. That's what custom instructions are for.
Custom instructions are notes you save once, and ChatGPT considers them in every new chat. According to OpenAI's help center, they're available on all plans, on web, desktop, iOS, and Android.
Changes apply to future conversations, not chats you've already started. There's also a character limit that depends on your plan, so keep it tight.
Here's an example you can adapt:
I write a marketing blog for small business owners. Write in plain, friendly English with short paragraphs. Use bullet points when listing more than 3 things. If my request is unclear, ask me up to 3 questions before answering. Never invent statistics. If you're not sure about a fact, say so.
Think of it as a permanent role and constraints prompt. Your normal prompts then only need the task itself.
Troubleshooting: Why Your Prompt Isn't Working
Before you rewrite the whole prompt, run through this quick checklist. It usually shows me what's missing in under a minute.
Quick Diagnostic Checklist
Problem: The Answer Is Too Long or Too Short
This happens when ChatGPT has to guess how much detail you want.
Problem: The Answer Is Off-Topic
Off-topic answers usually come from a vague goal or no boundaries.
Problem: The Answer Is Too Generic
Generic writing means the prompt has no clear reader, no constraints, or nothing to copy. This is where few-shot prompting shines.
Problem: The Answer Has Errors or Made-Up Details
Errors show up when ChatGPT has to fill gaps and nobody asked it to check its work. RAG-style and self-ask prompts help a lot here.
Problem: ChatGPT Ignores Your Instructions
When instructions are buried or fight each other, the model follows the wrong one.
Prompt Repair Template
When a prompt still isn't working, I paste it into this and fill in the blanks.
Task: [WHAT YOU WANT PRODUCED] Goal: [THE OUTCOME FOR THE READER] Audience: [WHO IT'S FOR + THEIR CONTEXT] Inputs: [FACTS, NOTES, LINKS, REQUIREMENTS] Output format: [STRUCTURE, HEADINGS, BULLETS, TABLES] Length: [WORD COUNT OR SECTION LIMITS] Tone/style: [TONE + WHAT THAT MEANS] Must include: [KEY POINTS] Must avoid: [TOPICS, CLAIMS, PHRASES] Assumptions: List your assumptions first. Ask up to [NUMBER] clarifying questions if needed. Self-check: Verify against my inputs, flag anything uncertain, and confirm every constraint is met.
Conclusion
You now know the 12 types of ChatGPT prompts and how I write each one. Focus on clarity, specificity, and context, and ChatGPT will give you answers you can actually use.
Pick one type you haven't tried yet and test it on a real task this week. That's honestly the fastest way to get better at this.
And if you just want to have some fun while you practice, try a few of my fun prompts for ChatGPT and see which type each one uses.
Frequently Asked Questions
What are the main types of ChatGPT prompts?
The ones I use most are zero-shot, few-shot, chain-of-thought, self-consistency, least-to-most, self-ask, ReAct, tree-of-thoughts, RAG-style, self-refine, structured output, and program-aided (PAL) prompts. Each one fits a different kind of task.
How long should a ChatGPT prompt be?
As long as it needs to be to cover the task, the context that changes the answer, and your constraints. Clarity beats length every time, and always say what output format you want.
What's the best structure for a prompt?
Goal, context, audience, constraints, and output format, plus an example if you have one. That simple structure works as a reusable template for most tasks.
How do I stop ChatGPT from making things up?
Ask it to state its assumptions, flag anything it is unsure about, and answer only from the sources you give it (RAG-style prompting). Then check the key facts yourself before you use them.
Where do I set ChatGPT custom instructions?
On web and desktop, open Settings and go to Personalization. On iOS and Android, open Settings and tap Customize ChatGPT. Your instructions apply to new chats from then on.
