ChatGPT Prompt Engineering Guide: 20 Copy-Ready Templates

Updated: 2026-08-12

Overview

Prompt engineering is specification writing. A strong prompt supplies the outcome, evidence, constraints, output contract, and quality test. It does not need theatrical role-play or magic phrases. Start short, inspect the failure, and add only the missing instruction.

The six-line prompt brief

Outcome: [decision or artifact]
Audience: [who will use it]
Evidence: [paste or attach source material]
Constraints: [scope, tone, length, prohibited claims]
Output: [headings, fields, or schema]
Quality test: [facts to check, examples to pass]
If evidence is missing, label it UNKNOWN and ask before assuming.

Templates 1–5: writing

1. Rewrite: “Rewrite for [reader] to achieve [action]. Preserve facts and numbers. Return the draft plus a list of meaning changes.”
2. Email: “Draft a [tone] email requesting [action] by [date]. Do not imply approval. Under 120 words.”
3. Outline: “Build a non-overlapping outline answering [question]. For each section state reader takeaway and evidence needed.”
4. Edit: “Mark unclear, unsupported, repetitive, and off-tone passages. Explain each edit before rewriting.”
5. Translate: “Translate to [locale] for [audience]. Preserve names, units, links, and formatting; list ambiguous terms.”

Templates 6–10: analysis

6. Compare: “Compare [options] on [weighted criteria]. Separate evidence from judgment and name missing data.”
7. Extract: “Using only the source, return JSON matching [schema]. Use null, never invention, for absent fields.”
8. Summarize: “Write decision, evidence, risks, and open questions. Preserve all dates and quantities.”
9. Research map: “Create claims to verify, best source type, search query, and evidence threshold. Do not answer yet.”
10. Critique: “Attack this proposal as finance, operations, customer, and security reviewers; rank issues by impact.”

Templates 11–15: learning and meetings

11. Tutor: “Diagnose my level with three questions, teach one concept, then give a problem without revealing the answer.”
12. Quiz: “Create 8 questions from these notes: recall 2, application 4, transfer 2. Provide a separate key.”
13. Explain: “Explain [concept] with a concrete mechanism, one analogy and its limit, then a check question.”
14. Minutes: “Separate decisions, action items, questions, and suggestions. Every decision needs a transcript quote.”
15. Interview prep: “Ask one [role] interview question at a time. Score my answer against this rubric and request a better second attempt.”

Templates 16–20: code and operations

16. Debug: “Given versions, expected/actual behavior, error, and minimal code, rank hypotheses; propose the smallest diagnostic test first.”
17. Tests: “Create tests from these requirements, covering happy path, boundary, invalid input, and regression. Do not implement.”
18. SQL review: “Review this query for correctness, null behavior, duplicates, performance, and data leakage; return a safer version.”
19. SOP: “Turn the process into trigger, prerequisites, numbered actions, decision points, rollback, owner, and evidence of completion.”
20. Plan: “Break the outcome into dependencies and milestones. Each task needs owner, acceptance test, risk, and rollback.”

Repair a weak answer

Do not start over with “be more detailed.” Identify the defect: missing evidence, wrong audience, mixed categories, unsupported claims, unusable format, or no acceptance test. Then issue a patch instruction such as: “Keep sections 1 and 3 unchanged. Rebuild section 2 using only the attached policy; cite paragraph numbers; mark conflicts rather than resolving them.”

Evaluate prompts instead of collecting them

Keep five representative inputs and an expected-result rubric. When a prompt changes, rerun all five and score accuracy, completeness, format validity, unsupported claims, and review time. In API use, version the prompt with model and parameter settings on OpenAI Platform. A shorter prompt that passes the test is better than a complicated one that merely sounds expert.

Before you trust the result

• Confirm names, dates, numbers, links, and quoted text against the source.
• Treat plan limits, pricing, model IDs, and regional availability as time-sensitive.
• Remove passwords, API keys, personal identifiers, and confidential business data.
• Test the output in the real destination before publishing or automating it.

Questions people ask

Which plan includes every feature?

No static answer stays accurate; check the current product and account page before purchasing.

Can I use outputs without review?

No. Verify facts, rights, confidential data, and task-specific acceptance criteria.

Official and useful links

• ChatGPT
• ChatGPT web entry
• OpenAI Platform

Continue learning

• What is ChatGPT?
• ChatGPT vs Claude vs Gemini