How to write better prompts for AI chat apps, with examples

Most disappointing AI answers start with a prompt that would confuse a human colleague too: "write something about our product" or "fix this" with no code attached. The model fills every gap with a guess, and the result feels generic because it is.

The good news is that prompting is not a secret skill. A short, consistent structure improves answers from every model, whether you are typing on a phone or dictating while walking. This guide gives you that structure, a few real examples, and the follow-up habits that matter more than any single phrasing.

Start with context, not the question

The model knows nothing about you, your audience or your project unless you say so. Two sentences of background change everything: who you are, who will read the result and what has already been decided.

Compare "Write a product description" with "I sell handmade ceramic mugs online to gift buyers. Write a product description for a 350 ml speckled mug; the tone is warm, not salesy." The second prompt is only one line longer and gets a usable draft on the first try.

If you repeat the same background in every chat, let the app keep it. Cortex AI's Smart Memory remembers your preferences and context across conversations, so you can stop re-explaining that you write in British English or that your code base uses TypeScript.

The four-part prompt that works everywhere

1. Context. Who you are and what the situation is. "I am a nurse preparing a short talk for new parents."
2. Task. One clear verb. "Explain how to read a baby's temperature" is better than "help me with the talk".
3. Constraints. Length, audience, what to avoid. "Under 200 words, no medical jargon, do not mention brand names."
4. Format. The shape of the output. "Three short paragraphs followed by a checklist of five items."
5. Put the four parts together and you have a complete prompt: "I am a nurse preparing a short talk for new parents. Explain how to read a baby's temperature. Under 200 words, no medical jargon, no brand names. Three short paragraphs followed by a five-item checklist."

Use the tools around the prompt

• Attach the source instead of describing it. For a contract, a report or a screenshot of an error, upload the PDF or image and ask about it directly. Cortex AI reads attached images and PDFs with full context, so "summarise section 3 of this file" works far better than retyping it.
• Use a search-capable model for anything time-sensitive. Models answer from training data that has a cut-off date. If the question is about prices, news, releases or schedules, pick a model with web search (when your plan includes it) so the answer is backed by live data, and check the sources it cites.
• Start from a template when you are stuck. Cortex AI ships more than 100 ready-to-use prompt templates for coding, writing, translation and analysis. Fill in the blanks, then edit the result to match your situation.
• Set the tone once. Instead of adding "be professional" to every prompt, adjust the tone settings: style, friendliness and enthusiasm can be set to professional, casual, academic or creative and apply to the whole conversation.
• Dictate long prompts. Speaking a detailed brief is often faster than typing it on a phone. Voice input with continuous dictation lets you talk through the context and task, then tidy the text before sending.

Refine instead of starting over

The first answer is a draft, not a verdict. Reply with a specific correction ("shorter, and keep the second paragraph"), ask for alternatives ("give me three versions of the opening line") or ask the model to critique its own answer before you do. Each follow-up carries the whole conversation, so you do not need to repeat the context.

If a model keeps missing the point, change the model rather than the prompt. In Cortex AI you can switch between GPT-4, Claude, Gemini, DeepSeek, Grok and Qwen mid-conversation with one tap, or send the same prompt to two models at once and keep the better answer. Often the prompt was fine and the fit was wrong.

Mistakes that quietly ruin answers

• Several questions in one message. Ask them one at a time or number them, otherwise the model answers the easiest and skims the rest.
• Asking for "the best" without criteria. Best for whom, by what measure? State the criteria and you get a ranking you can trust.
• Pasting sensitive data you do not need to share. Prompts and attachments are sent to the AI provider that answers them. Remove names, account numbers and anything private that the task does not require.
• Treating the answer as verified. AI output can be wrong or out of date. For facts, figures and code that will ship, check against a source you trust before you rely on it.