12 ChatGPT Prompts That Actually Work: Practical Templates

Reviewed: August 12, 2026. A useful ChatGPT prompt is not a magic sentence. It is a small specification: the task, relevant context, constraints, expected output, and a way to check the result. The 12 reusable prompts below are designed for real work, not for filling a “500 prompts” list.

Replace every item in square brackets before using a template. Do not paste passwords, private customer data, unpublished source code, medical records, or other confidential material into a tool unless your organization has approved that use.

What makes a ChatGPT prompt work?

OpenAI’s current ChatGPT guidance recommends clear, specific requests with enough context, followed by iterative refinement. That does not mean every prompt must be long. It means every detail should reduce a real ambiguity.

If a competent colleague could not complete the task from your brief, the model probably cannot either. Add the missing evidence instead of adding dramatic phrases such as “be a world-class expert.”

12 ChatGPT prompts you can adapt

1. Turn a vague request into a clear brief

I need to complete this task: [task].

Before proposing a solution, ask me up to five questions that would materially change the result. Do not ask for information that is merely nice to have. After I answer, restate the agreed objective, constraints, deliverable and definition of done.

Use this when you know the goal but have not defined scope. The “materially change” condition prevents a long interview about minor preferences.

2. Summarize a document for a decision

Summarize the document between <document> tags for [decision-maker].

Return:
1. the decision required;
2. the three facts most relevant to that decision;
3. unresolved risks or missing evidence;
4. the next action and owner.

Use only the supplied document. If a point is not supported, label it “not stated.”

<document>
[paste text]
</document>

This is better than “summarize this” because it defines the reader and purpose. Treat copied text as data, not as trusted instructions.

3. Rewrite an email without changing the facts

Rewrite the email below for [recipient]. Keep every factual claim and commitment unchanged.

Goal: [what the recipient should understand or do]
Tone: calm, direct and respectful
Length: 90–130 words
Avoid: blame, filler, exclamation marks and promises not present in the original
Output: subject line followed by the email

Original email:
---
[email]
---

Always compare the revision with the original before sending. Language models can silently strengthen a promise or alter a date.

4. Extract structured data

Extract the following fields from each record:
- customer_name
- invoice_number
- invoice_date in YYYY-MM-DD format
- total_amount as a number
- currency as a three-letter code

Return a JSON array only. Use null when a value is missing; do not infer it. Add a final field named source_quote containing the exact short text that supports each extracted record.

Records:
---
[records]
---

A defined schema makes extraction testable. The source quote gives a reviewer a quick route back to the evidence.

5. Explain a technical concept at two levels

Explain [concept] twice:

A. Beginner explanation: 150 words, one concrete analogy, no unexplained jargon.
B. Practitioner explanation: include the mechanism, one realistic use, one limitation and one common failure mode.

Finish with three questions I should be able to answer if I understood it. Do not provide the answers until I attempt them.

The second explanation stops an analogy from replacing the real mechanism. The questions turn passive reading into a checkpoint.

6. Review code without inventing execution results

Review the code below for correctness, security, maintainability and edge cases.

For each issue provide:
- severity: critical, high, medium or low;
- exact function or line fragment;
- why it can fail;
- smallest safe fix;
- a test that would expose the problem.

Do not claim you ran the code. Separate confirmed defects from possible concerns.

Code:
```[language]
[code]
```

A useful review distinguishes evidence from suspicion. Run the proposed tests in your own environment before accepting a fix.

7. Create a study plan around a project

Create a [number]-week beginner study plan for [skill]. I can study [hours] hours per week and can use only free tools.

My starting knowledge: [knowledge]
Portfolio goal: [project]

For each week include one learning objective, one practice exercise, one visible project milestone and one pass/fail checkpoint. Put prerequisites before dependent topics. Do not include paid courses or guaranteed career claims.

A project milestone is stronger than “watch three videos.” It produces evidence the learner can inspect and improve.

8. Compare options using stated criteria

Compare [option A], [option B] and [option C] for this situation: [context].

Criteria and weights:
- [criterion]: [weight]%
- [criterion]: [weight]%
- [criterion]: [weight]%

First identify any facts that need current verification. Then produce a comparison table, explain each score in one sentence, and show how the recommendation changes if the most important assumption is wrong. Do not invent prices or features.

Weights make the recommendation traceable. For purchases or current software, verify specifications and prices at the source.

9. Diagnose a problem before proposing fixes

Help me diagnose this problem: [symptom].

Known facts:
- [fact]
- [fact]

Recent change: [change or none]
Environment: [system/version]

Rank the five most plausible causes. For each, give one safe read-only check and explain what a positive or negative result would mean. Do not suggest destructive changes until the cause is supported.

This format prevents random “try everything” troubleshooting. Perform one check at a time and record the result.

10. Build a content brief from search intent

Create a content brief for the query “[query].”

Audience: [audience]
Site expertise: [what the publisher can genuinely support]
Business goal: [goal]

Separate what requires live search research from what can be explained from stable knowledge. Propose one primary intent, an outline, questions the page must answer, original examples or assets to create, internal-link opportunities and claims that require primary sources. Do not invent search volume.

This prompt creates a research plan, not finished SEO copy. Current SERPs, cannibalization, and factual sources still require verification.

11. Critique a draft against a rubric

Evaluate the draft against this rubric: [rubric].

Return a table with criterion, evidence from the draft, score, and one specific revision. Quote no more than one short sentence as evidence. Then list the three revisions with the highest impact.

Do not rewrite the draft yet. If a criterion cannot be assessed from the text, mark it “insufficient evidence.”

Draft:
---
[draft]
---

Separating evaluation from rewriting makes it easier to judge whether the diagnosis is sound before accepting changes.

12. Produce options that differ meaningfully

Create three versions of [deliverable] for [audience]. All must preserve these facts: [facts].

Version A: safest and most conventional
Version B: concise and direct
Version C: more distinctive but still credible

After each version, explain its main trade-off in one sentence. Do not create superficial variants that only change synonyms.

Multiple options help you compare trade-offs. The axes must be explicit or the model may return three nearly identical drafts.

How to test whether a prompt is reliable

1. Prepare two or three representative inputs. Include one normal case and one awkward edge case.
2. Write a short rubric. State factual, formatting, safety, and completeness requirements.
3. Run the same prompt more than once. A single good output does not prove consistency.
4. Check claims against sources. A confident sentence is not evidence.
5. Change one instruction at a time. Otherwise you will not know which change helped.
6. Save failures. A small set of failed examples becomes a regression test for future revisions.

For the underlying method, see how to write AI prompts that get better answers. That page teaches how to design the brief; this page provides task-specific starting templates.

Prompting mistakes that waste time

• Requesting hidden reasoning: ask for a concise rationale, calculation, checklist, or cited evidence you can evaluate, not private chain-of-thought.
• Using vague quality words: replace “professional” or “engaging” with observable requirements.
• Combining unrelated jobs: split research, drafting, fact-checking, and final formatting when each needs a different review.
• Treating examples as facts: label placeholders and verify every real name, number, date, link, and quotation.
• Blindly copying prompt libraries: preserve the useful structure but replace the audience, evidence, limits, and success criteria.
• Trusting tool claims: confirm that an action, test, search, or file operation actually occurred.

Frequently asked questions

Do longer prompts always work better?

No. A prompt should be long enough to remove important ambiguity. Extra repetition can obscure priorities. Start with the task and essential context, then add constraints that change the answer.

Should I tell ChatGPT to act as an expert?

A role can focus attention, but it does not create credentials or guarantee accuracy. State the audience, decision, risk, and standards the response must meet. Verify high-stakes claims with qualified people and primary sources.

Can I use these prompts with other AI assistants?

Most patterns are model-agnostic because they clarify the work. Specific capabilities, context limits, privacy controls, and tool access differ by product, so check the provider’s current documentation.

What should I do when the first answer is wrong?

Name the exact failure, supply missing evidence, and request a targeted revision. Start a new conversation when earlier context is unrelated or is causing confusion, not merely because the first draft was imperfect.

Official OpenAI references

• Prompt engineering best practices for ChatGPT
• How to create a good prompt for an AI model
• Prompt engineering guidance for the OpenAI API

Written and reviewed by Muhammad Azhar. Examples are educational templates; review outputs before using them in production, legal, medical, financial, employment, or security decisions.