AI tools

How to Write AI Prompts That Work First Time

By Sauravjit Singh · Updated 2026-10-28 · 10 min read

Short answer: To write a good AI prompt, tell the tool five things: who it should act as, the exact task, the context only you know, the format you want back, and the limits (length, tone, what to avoid). Add an example if you have one. Then read the output critically and ask for one specific fix at a time.

Most weak AI output is not the tool's fault. It is a one-line request that leaves the model guessing about the reader, the purpose and the facts. A model fills those gaps with safe, generic text, which is why so much AI writing sounds the same.

This guide gives you one checklist that works in ChatGPT, Claude and Gemini alike, then shows it in action on three ordinary Indian work tasks: a GST reminder email, product copy for a boutique, and a job description. Each rewrite is critiqued line by line, so you can see why the second version works and spot the same problems in your own prompts.

The prompt checklist (use it in any tool)

A good prompt removes guesswork. The model has read a huge amount of text, but it knows nothing about your client, your shop or your company until you say so. In our agency work since 2013, the pattern has been the same with every new team member and now with every AI tool: the brief decides the quality of the first draft. A prompt is a brief. If you would not hand a one-line brief to a junior colleague, do not hand it to ChatGPT.

Three principles sit under the checklist. Specific beats long: fifty precise words outperform three hundred vague ones. Facts must come from you, because the model will invent plausible dates, amounts and policy names if you leave them out. And format is part of the task: say 'a table with three columns' or 'under 120 words' instead of hoping.

Use these six parts as a checklist, not a rigid template. Short tasks may need only three of them. A prompt framework is simply a reminder of what the model cannot know on its own.

PartWhat to writeExample phrase
1. RoleWho the model should act as, and for whomAct as an accounts executive at a small CA firm writing to a client
2. TaskOne clear verb and the deliverableWrite a polite payment reminder email
3. ContextFacts only you know: names, numbers, history, audienceInvoice no. 214, ₹18,500 incl. GST, due 10 days ago, client is usually prompt
4. FormatShape, length, structureSubject line plus body under 120 words
5. ConstraintsTone, words to avoid, what not to doWarm, not threatening; no legal language; do not invent bank details
6. Example or checkA sample you like, or ask it to flag gapsIf any detail is missing, ask me before writing

The sixth line is the most underrated. Asking the model to ask you questions first stops it from filling gaps with invented facts. It works the same way in all three major assistants; for which assistant suits which job, see our comparison of ChatGPT, Claude and Gemini.

In this guide: 28 related articles

Rewrite 1: a GST payment reminder email

Say you work at a small CA firm in Janakpuri and a client has not paid last quarter's GST filing fee.

Before: 'Write an email asking a client to pay their pending invoice.'

What comes back is usually a generic letter: 'Dear Valued Client, we hope this email finds you well', a vague mention of 'the outstanding amount', and a closing line about 'further action'. Line by line, the problems are clear. The greeting is impersonal. There is no invoice number or amount, so the client has to dig. 'Further action' sounds like a threat to a client who may simply have forgotten. And it is twice as long as it needs to be.

After: 'Act as an accounts executive at a small CA firm in Delhi writing to a long-standing client. Write a payment reminder email for invoice no. 214 dated 1 October for GSTR-1 and GSTR-3B filing for July–September, ₹18,500 including 18% GST, now 10 days overdue. The client, Mr. Arora, usually pays on time, so assume it slipped his mind. Give a subject line and a body under 120 words. Tone: warm and brief. Mention that payment details are in the attached invoice. Do not invent bank details, penalties or deadlines. If anything is missing, ask me first.'

  • Role and relationship set the tone: 'long-standing client' turns a demand into a nudge.
  • Invoice number, period and amount make the email actionable in one read.
  • 'Assume it slipped his mind' is context the model could never guess, and it changes every sentence.
  • 'Do not invent bank details' blocks the single most risky hallucination in a payment email.

Critique the second output too. A typical good draft still opens with a filler line like 'I hope you are doing well'. Your follow-up should be one fix: 'Cut the first line and put the amount in the subject.' One instruction per round is faster than rewriting the whole prompt.

Rewrite 2: product copy for a Rajouri Garden boutique

Say you run a boutique in Rajouri Garden selling hand-embroidered kurta sets, and you need Instagram and website copy for a new piece.

Before: 'Write a product description for a designer kurta.'

The typical output calls the kurta 'elegant', 'timeless' and 'perfect for every occasion'. Each of those words could describe any garment in India. There is no fabric, no work detail, no size or care information, and no reason to choose this kurta over the one in the next shop.

After: 'Act as a copywriter for a West Delhi boutique whose buyers are working women aged 25–40 shopping for festive and wedding-guest wear. Write a product description for a mint green chanderi kurta set with hand-done gota patti on the yoke, straight pants and an organza dupatta. Sizes S to XXL, dry clean only, ships from Delhi in 3–5 days. Give: a 6-word product title, a 60–80 word description that leads with how it looks and feels when worn, and 4 bullet points for fabric, work, fit and care. Avoid the words elegant, timeless and stunning. Indian English, no emoji.'

  • The buyer description (working women, festive and wedding-guest wear) decides which details the copy leads with.
  • Real fabric and work names (chanderi, gota patti, organza) make the copy specific and help it rank for the way people search.
  • Banning three overused adjectives forces the model to describe instead of praise.
  • Asking for title, description and bullets separately gives you pieces you can paste into Instagram, the website and a WhatsApp catalogue.

If the output still invents something, such as 'pure silk' when you said chanderi, correct it explicitly. Never publish product claims you have not checked against the actual garment.

Rewrite 3: an HR job description

Say you are an HR executive at a 40-person logistics firm in Gurugram hiring an operations coordinator.

Before: 'Write a job description for an operations coordinator.'

The usual result is a long list of responsibilities copied from the global average of every JD online: 'liaise with cross-functional stakeholders', 'drive operational excellence', and a requirement for '5+ years of experience' that you never asked for. Candidates skim it and learn nothing about the actual job.

After: 'Act as an HR executive at a 40-person logistics company in Gurugram. Write a job description for an Operations Coordinator. The person will track 25–30 daily shipments on our dispatch sheet, call transporters when a truck is late, update clients on WhatsApp, and prepare a daily delay report for the operations head. Requirements: graduate, 1–3 years in any coordination or back-office role, comfortable with Excel and spoken Hindi and English. Location and terms: Gurugram office, 6 days a week, salary as per industry norms. Format: a 2-line summary, 6 responsibility bullets, 5 requirement bullets, and a 'what a normal day looks like' paragraph of 60 words. Plain language, no corporate jargon. Do not add benefits we have not listed.'

  • Concrete daily tasks (shipments, transporters, WhatsApp updates) replace jargon and attract candidates who can actually do the work.
  • The experience range is stated, so the model stops inflating requirements.
  • 'Do not add benefits we have not listed' prevents promises HR cannot keep, such as health insurance or hybrid work.
  • The 'normal day' paragraph is a structure most templates miss, and candidates read it first.

How do you fix a bad AI answer?

Do not start again. Tell the model exactly what is wrong, one issue at a time, and ask for a revision. Iteration is where most of the quality comes from.

  1. Read the output against your checklist. Did it follow the format, length and tone?
  2. Mark every fact it added that you did not give it. Delete or verify each one.
  3. Give one correction: 'Shorten paragraph two to two sentences' or 'Replace the closing with a question'.
  4. If it keeps drifting, paste a sample of writing you like and say 'match this tone and sentence length'.
  5. Once the result is right, ask: 'Rewrite my original prompt so it would have produced this first time.' Save that prompt.

That last step builds your personal prompt library. Saved prompts are also what you put into ChatGPT custom instructions or a reusable ChatGPT Project, so you stop retyping context.

Mistakes beginners make with prompts

A common mistake we see from beginners is treating the AI like a search box: three keywords and enter. Other repeat offenders:

  • Asking for 'the best' or 'a viral' version. These words carry no instruction; describe the outcome instead.
  • Stacking five tasks in one prompt. Split research, outline and draft into separate turns.
  • Trusting numbers, laws and dates. Always run AI facts through a quick fact-checking routine for AI answers, especially GST rates, deadlines and legal sections.
  • Pasting confidential client data without thinking. Read what is safe to upload to ChatGPT before sharing PAN numbers, salaries or contracts.
  • Copying 'mega prompts' from social media. They are long, not specific, and rarely fit your task.
  • Writing prompts in a language you do not think in. If Hindi is easier, prompting ChatGPT in Hindi works well for drafting, and you can ask for the final output in English.

Does the same prompt work in ChatGPT, Claude and Gemini?

Yes, mostly. The checklist above works across all three because they are all trained to follow clear instructions. Differences show up in default length, tone and how readily each tool asks clarifying questions, not in the basics.

A few patterns we see when the same prompt runs in all three: one tool may return a longer draft with headings you did not ask for, another may ask a clarifying question first, and another may stay closer to your exact wording. None of this changes the method. If a tool overshoots on length, the fix is the same everywhere: state the word limit and the number of paragraphs. If it skips a fact, move that fact to the start of the prompt.

Save each finished prompt in one place, such as a Google Doc titled by task ('GST reminder', 'product copy', 'JD'), with the date and the tool it worked best in. After a month you will have a dozen prompts that cover most of your week.

In practice, run your saved prompt in a second tool when the first output feels flat. If you are unsure which assistant to start with, our existing post on which AI tool to use for what maps tools to tasks. For specific tools, see the beginner walkthroughs for using ChatGPT step by step, getting started with Claude and Gemini inside Gmail and Docs.

In this guide: every AI tools article

Use the checklist with these focused guides. Each one covers a single tool or task.

Prompting is the core skill in the Applied AI Professional programme at Spettro School, where every module uses real work tasks rather than toy examples. If you want structured practice, our no-code AI course covers these tools over 8 weeks.

Frequently asked questions

What is the best prompt framework for beginners?

Any framework that covers role, task, context, format and constraints will do. The six-part checklist in this guide is a good start because it also asks the model to flag missing information before it writes.

How long should an AI prompt be?

Long enough to include every fact the model cannot guess, and no longer. For most office tasks that is 50 to 150 words. A long prompt full of adjectives is worse than a short one full of specifics.

Can I write prompts in Hindi or Hinglish?

Yes. ChatGPT, Claude and Gemini all understand Hindi and Hinglish prompts. You can prompt in Hindi and ask for the answer in English, which many people find faster for drafting.

Do I need to learn prompt engineering to use ChatGPT well?

No. For everyday work you need clear briefing and critical reading, not technical tricks. The deeper techniques matter mainly for people building AI products or automations.

Why does ChatGPT give different answers to the same prompt?

These models generate text with some randomness, so each run varies. Tight format and constraint lines reduce the variation, and a sample of the output you want reduces it further.

Want to learn this with a trainer?

Leave your details and our counsellor will call you within one working day with batch dates, fees and the right course for you.

  • Free counselling call, no obligation to join
  • Weekday and weekend batches, online or offline
  • Small batches of 15 to 20 students

Or chat on WhatsApp

Enter a 10-digit mobile number.

We use your details only to contact you about Spettro School courses.

Chat on WhatsApp