Prompting ChatGPT in ways that actually work
Prompt patterns that work, anti-patterns that fail, and an honest iteration process.
ChatGPT guide · as of July 10, 2026 · 4 minutes read · details change — confirm current specs on chatgpt.com
Good prompting is not about secret phrases. It is about giving ChatGPT enough context to do a defined job, then checking and improving the result.
The best prompt cannot supply facts you do not have, make an unreliable source reliable, or turn a probabilistic model into deterministic software.
The five-part prompt
A practical prompt often contains:
- Context: What situation is this for?
- Task: What should ChatGPT do?
- Constraints: What must it preserve, avoid, or limit?
- Format: What should the answer look like?
- Verification: What should it flag or check?
Example:
I am comparing two contractor proposals for the same kitchen. Create a table of scope, exclusions, allowances, schedule, and payment terms. Quote the exact proposal language for every difference. Do not recommend a contractor yet. Flag anything that cannot be compared.
| Prompt element | Why it helps |
|---|---|
| Context | Prevents generic advice |
| Clear task | Gives the response a finish line |
| Constraints | Protects facts, tone, and scope |
| Output format | Makes the result easier to inspect |
| Verification request | Encourages visible uncertainty |
Ask for questions before answers
When important context is missing, tell ChatGPT not to guess:
Ask up to five questions, one at a time, before recommending anything.
This works well for purchases, travel, career decisions, technical troubleshooting, and writing for an unfamiliar audience.
It works less well when you already supplied the relevant details. Endless questioning can become avoidance. Set a limit.
Break large jobs into stages
Do not ask for research, analysis, drafting, fact-checking, and final formatting in one breath.
A better sequence is:
- Define the goal and audience.
- Gather or upload source material.
- Ask for an outline or comparison.
- Correct misunderstandings.
- Draft one section.
- Review against a checklist.
- Produce the final format.
This makes errors easier to spot and reduces the chance that a polished final answer hides a bad assumption.
Give examples carefully
Examples are powerful, but ChatGPT may copy their surface style too literally.
| Instruction | Likely result |
|---|---|
Make it professional | Generic business language |
Use short sentences, plain words, and no hype | More controllable style |
Match this sample's structure, not its facts | Better separation of form and content |
Preserve every number and proper noun | Lower risk of factual drift |
Explain all changes | Easier review |
For repeated work, provide one good example and one bad example with a short explanation of the difference.
Useful prompting patterns
The editor
Rewrite this for clarity. Preserve all factual claims. List any sentence whose meaning is ambiguous before changing it.
The interviewer
Help me choose between these options. Ask one question at a time, then show how each answer changes the recommendation.
The critic
Review this plan for missing assumptions, failure points, and costs. Do not rewrite it until the critique is complete.
The source-bound researcher
Answer only from the attached files. Cite the file and section for each claim. Say "not found" when the files do not support an answer.
The checker
Compare the draft against this checklist. Make a pass-fail table and quote the exact sentence responsible for each result.
Anti-patterns
Treating confidence as accuracy
“Are you sure?” often produces a more confident restatement. Ask for evidence, assumptions, and competing explanations instead.
Huge role prompts
“You are the world's greatest lawyer, doctor, programmer, and strategist” adds theater, not authority. Supply the jurisdiction, source, goal, and constraints.
Asking for hidden reasoning
You do not need a private internal monologue. Ask for a concise explanation, calculations, cited evidence, and a verification plan.
One-shot perfection
The first answer is a draft. Tell ChatGPT what failed and request a targeted revision.
Vague freshness
For current facts, say Search the web, use sources dated within the last 30 days where appropriate, and distinguish event date from publication date. Then inspect the sources.
Make corrections durable inside the chat
When the conversation drifts, do not keep patching individual sentences. Restate the working brief:
Use this as the controlling brief from now on: audience, goal, constraints, approved facts, and output format.
For very long work, start a new chat with a clean summary. Old conversations can accumulate contradictions.
When this is the wrong tool
Prompting is the wrong solution when the underlying task needs exact rules, guaranteed calculations, a database lookup, legal authority, or professional judgment.
A better prompt cannot make ChatGPT safely diagnose an emergency, file your taxes, verify identity, or control a dangerous device. Use software and people designed for those jobs.
The real skill is not producing a clever sentence. It is designing a process in which a plausible mistake is visible before it matters.
