Prompting guides
How to write better AI prompts
A good AI prompt states the goal, gives the facts the model cannot guess, and says what a good answer looks like. The same six parts work in ChatGPT, Claude, Gemini and Perplexity, and each tool has a few small differences that this guide maps.
macOS 13+·Windows 10+·Free for 3 days, no card
Updated
>_The basics
The short answer: a prompt in six parts
A good prompt answers six questions, and a one-line prompt answers almost none of them.
- Goal: what you want, and why it matters.
- Facts: the names, dates, numbers and documents the model cannot guess.
- Steps: what the model should do, in order.
- Limits: the tone, the audience, what to avoid, and a permission to say "I am not sure".
- Format: the shape and length of the answer.
- A great answer: what good looks like, so the model can aim at it.
No guide uses these six labels. They are our grouping of advice that the guides below share. Anthropic describes Claude as a new colleague who lacks context, and its test is to show your prompt to a person who knows nothing about your work. If that person would be confused, Claude will be too. We find the test useful in any tool.
You do not need all six every time. Google's Workspace guide says you do not have to use all four of its parts. Microsoft's guide says a clear goal is the only required part. A quick question needs the goal and little more. A task for a client needs all six.
The problem
Why AI answers come out generic
The model has only the words you give it, so a short prompt gets the average answer.
A one-line prompt tells the model nothing about your customer, your deadline or your voice. It fills the gaps with the most likely guess. That guess reads as generic.
Microsoft's help page gives the plain reason for detail: Copilot has only your current conversation to work from. Three mistakes cause most generic answers:
- Vague words. "Make it better" and "something short" leave the model guessing. Guides from Google, GitHub and Perplexity all tell you to cut vague or ambiguous wording.
- Instructions that fight each other. OpenAI's GPT-5 recipe, an archived page, says contradictory instructions make the model spend effort trying to reconcile them.
- No limits and no way out. Without a length, an audience or a permission to say "I am not sure", the model has to guess at those too. Anthropic's blog says this permission reduces made-up answers.
Typing six parts every time is slow, so Scrubbr writes them for you, from your own email, chats and notes, in any box where you type.
Shared habits
Where the guides agree
Guides from Anthropic, Google, Microsoft, GitHub and Perplexity give advice of this shape, and OpenAI and Cursor back some of it.
Each row lists some of the guides that say it. We checked every row against the guides on 2026-11-03. OpenAI's rows come from its developer guides, not from its help centre.
| Habit | Who says it |
|---|---|
| 1. State the goal and the task in plain, specific words. | Anthropic. Google. Microsoft. Perplexity. GitHub. |
| 2. Give context or source material the model cannot guess. | Anthropic. Google. Microsoft. Perplexity. Claude Code. |
| 3. Say the audience, tone, format and length. | Anthropic. Google. Perplexity. Microsoft. |
| 4. Break big jobs into steps or separate prompts. | Anthropic. Google. Perplexity. GitHub. |
| 5. Show an example of what good looks like. | OpenAI. Anthropic. Google. GitHub. |
| 6. Allow "I don't know", then check the answer yourself. | Anthropic. Perplexity. Microsoft. GitHub. |
| 7. Iterate with follow-ups. Two coding tools say the opposite for a bad result: Cursor says revert, refine the plan and rerun. Claude Code says clear the chat and write a better prompt after two failed corrections. | Google. Perplexity. Microsoft. Cursor. Claude Code. |
| 8. Separate the parts clearly, with headings or tags. | OpenAI. Anthropic. Google. |
| 9. Remove vague or conflicting instructions. | OpenAI's archived recipe. GitHub. Google. Cursor. |
| 10. In coding tools: read the code first, plan first, and give a pass or fail check. | Claude Code. Cursor. GitHub. |
How the six parts map: Goal is habit 1. Facts is habit 2. Steps is habit 4. Limits is habits 3 and 6. Format is habit 3. A great answer is habit 5, plus the check in habit 6.
Tool by tool
Where the tools differ
The core is the same everywhere, and each tool adds a few habits of its own.
| Tool | What is different | Full guide |
|---|---|---|
| ChatGPT | OpenAI's developer guide keeps the rules apart from the inputs, and suggests a layout of identity, instructions, examples, then context. | How to prompt ChatGPT |
| Claude | Examples work well. Put long documents first and your question last. | How to prompt Claude |
| Gemini | Google's Workspace guide names four parts: Persona, Task, Context and Format. | How to prompt Gemini |
| Perplexity | Perplexity's help guide says to state your intent, add detail and ask for one thing at a time. | How to prompt Perplexity |
| Microsoft Copilot | Four parts: Goal, Context, Source, Expectations. Only the goal is required. | How to prompt Microsoft Copilot |
| Claude Code | Tell it what it cannot see, name the file and the symptom, and give it a check to run. | How to write a prompt for Claude Code |
| Cursor | Plan before you code, keep rules short, and start a new conversation for each task. | Cursor prompting tips |
| GitHub Copilot in VS Code | Goal first, then requirements with the most important first. Standing rules live in instruction files. | GitHub Copilot prompts in VS Code |
Worked example
From one line to a six-part prompt
Dana needs to follow up with Birchfield Foods after a pilot call, and she says it the way she thinks it.
In ChatGPT
You say
optionwrite a follow up to the Birchfield thing, they liked the pilot but they are nervous about timing, keep it short and friendly and get a call booked this week
Scrubbr finds
Gmail · Birchfield Foods thread
Granola · Birchfield pilot call
ChatGPT · Prompt
Goal: Write a follow-up email to Chloe Barrett at Birchfield Foods that books a decision call this week. The pilot ends on Nov 14, and a gap would stall the rollout.
Facts: Chloe is the operations lead. On the Oct 28 call she said the 12-seat pilot went well, but her team is busy until mid-November and she worries about the start date. Her Nov 1 email asks for the price per seat and a written start date. Our price is $18 per seat per month. We can start on Dec 2. Evan Foster signs for Birchfield.
Steps: 1. Thank her for the pilot feedback in one line. 2. Answer her two questions: price and start date. 3. Say the Dec 2 start gives her team time to finish the busy weeks. 4. Ask for a 20-minute call and offer Thursday or Friday.
Limits: Friendly and plain, as Dana writes to Chloe. No pressure and no discount. Do not invent a date or a price. If something is missing, flag it for Dana in a separate note.
Format: A plain email with a subject line. Under 110 words.
A great answer: Chloe can reply "Thursday works" without asking another question.
Demo data. The people, company and facts are invented.
Questions
Prompting questions
How do I write better AI prompts?
State the goal, give the facts the model cannot guess, and say the steps, the limits and the format. Then say what a good answer looks like. Guides from Anthropic, Google and Microsoft give advice of this shape.
Why are my AI answers so generic?
A short prompt gives the model no facts about your situation, so it falls back on the average answer. Anthropic's guide pictures Claude as a capable new employee who lacks your business context.
Do prompts work the same in ChatGPT, Claude and Gemini?
The core habits are the same. The differences are small. Anthropic's guide puts long documents first and the question last, and Google's Workspace guide names four parts: Persona, Task, Context and Format.
Do I need all six parts every time?
No. Google's guide says you do not need all four of its parts, and Microsoft says a clear goal is the only required part. Add the other parts when the task is for someone else or the cost of a miss is high.
Sources
Where these facts come from
Sources checked on 2026-11-03.
- Prompt engineering (OpenAI)
- GPT-5 prompting guide (OpenAI Cookbook, archived)
- Prompting best practices (Anthropic)
- Best practices for prompt engineering (Anthropic blog)
- Best practices for Claude Code
- Prompt design strategies (Gemini API)
- Writing effective prompts (Google Workspace)
- Tips to write prompts for Gemini (Google Workspace Learning Center)
- Prompt Guide (Perplexity Agent API)
- Tips for Getting Better Answers from Perplexity
- Practical Tips for Using Perplexity
- Best practices for coding with agents (Cursor)
- Rules (Cursor)
- Prompt engineering for GitHub Copilot
- Customize agents for your workflow (VS Code)
- Get started writing prompts in Microsoft Copilot
- Write a great prompt in Microsoft Copilot
- Get better Copilot responses with great prompting
Scrubbr writes this prompt for you. Hold Option and say one line.
Try Scrubbr free for 3 days
Download Scrubbr, connect your tools, and give your first prompt in two minutes.
macOS 13+·Windows 10+·Free for 3 days, no card