A prompt is the instruction, question or context you give an AI tool. It is the only thing the model has to work with, and it determines almost everything about the quality of what comes back.
Every interaction with ChatGPT, Claude, Gemini or an image generator starts with one. It can be three words or three paragraphs. The difference in output between those two is usually larger than the difference between two competing models.
Why the wording matters so much
A large language model does not infer what you meant. It continues from what you wrote. Your prompt sets the entire statistical context for what follows, so every detail you leave out is a detail the model fills in with whatever is most generic.
Ask for write an email and you get a bland template, because the model had nothing to narrow it down. Ask for write a 120 word email to a client in Mumbai, politely following up on an invoice that is 20 days overdue, warm but firm, no apology at the start and you get something you can send with light editing.
Nothing about the model changed. You simply removed the ambiguity it was going to resolve badly.
The parts of a good prompt
Not every prompt needs all of these, but most weak prompts are missing several.
- Role. Who the model should act as. "You are an experienced Indian tax consultant explaining to a first-time filer."
- Task. The specific thing to do, stated as a verb. Summarise, rewrite, compare, draft, critique, explain.
- Context. Background it cannot know. Your audience, your constraints, what has already been tried.
- Format. Length, structure, tone. Bullet points or paragraphs. 200 words or 800.
- Examples. One or two samples of the output style you want. This is the single most effective addition for a specific style.
- Constraints. What to avoid. Jargon, assumptions, or making things up.
Weak and strong, side by side
| Weak prompt | Strong prompt |
|---|---|
| Summarise this | Summarise this report in three sentences for a busy manager, focusing only on the financial risks |
| Write a blog post about SIPs | Write a 700 word explainer on SIPs for a 25 year old first-time investor in India, using rupee examples, no jargon, and a short section on what people get wrong |
| Is this a good investment | Explain the main risks and trade-offs of this investment in plain English, list what information is missing before anyone could judge it, and do not give a recommendation |
| Fix my code | This Python function should return the EMI for a loan but returns a negative number for long tenures. Here is the code and a failing example. Explain what is wrong and show the corrected version |
Techniques worth knowing
Iterate rather than perfect. Send a reasonable prompt, read the output, then say what to change. "Shorter." "More formal." "Use an Indian example." Conversation almost always beats trying to write the ideal prompt in one go.
Ask it to think step by step for reasoning tasks. It genuinely improves accuracy on multi-step problems, because the model produces intermediate reasoning it can build on rather than jumping to an answer.
Give it the source. Pasting the actual document and asking it to work only from that reduces invention substantially, though it does not eliminate it. See hallucination.
Ask what is missing. "What else would you need to know to answer this well?" often surfaces assumptions you should have specified.
Split large tasks. One prompt for an outline, another for each section, produces better results than one prompt for an entire document.
Practical notes for users in India
Language. English generally gives the strongest results, Hindi is handled reasonably, and other Indian languages vary. Writing in transliterated Hinglish often works better than people expect, and asking for the output in your preferred language while prompting in English is a useful hybrid.
Be specific about Indian context. Models default to American assumptions. Saying "in India", "in rupees", "under Indian tax law" or "for the Indian market" changes the output substantially and is worth doing every time.
Do not put confidential data in the prompt. Everything you type may be retained and used for training by free consumer tools. Keep PAN, Aadhaar, bank statements, salary details and client information out of them.
Verify anything factual. A well-crafted prompt improves clarity and relevance. It does not make the model's facts correct, and it will not stop it inventing a section number or a rate.
Getting consistently good at this is a learnable skill in its own right. See prompt engineering, and try it in our AI assistant.