A Handbook for Using Generative AI
(A Reference Toolkit)
Disclaimers
This handbook has been developed by YASHADA's Centre of Excellence in Artificial Intelligence as a learning and capacity-building resource for officers, trainers, and other learners, to build awareness and support informed, safe, ethical, and responsible use of Generative AI.
A significant part of this handbook has been designed and developed with the assistance of Generative AI tools, including ChatGPT, Claude, and Gemini, with human verification, due diligence, and human-in-the-loop (HITL) practices integrated throughout its development and validation.
Generative AI tools can produce plausible but inaccurate, outdated, or fabricated outputs. All AI-generated information, examples, and content in this handbook, and any output users generate using AI tools, must be independently verified before use or reliance.
Do not input classified, official, confidential, or personally identifiable information into public or commercial AI tools. Users must follow applicable government data classification, IT security, and confidentiality guidelines when using any AI tool, and remain responsible for safeguarding official and personal information at all times.
Information in this version has been reviewed for relevance and accuracy as of 19 September 2026. AI technologies, tools, and the regulatory environment evolve rapidly; readers should verify current information before relying on specific features, procedures, or requirements. This handbook is intended to be reviewed and updated biannually.
This handbook is a learning resource and does not constitute legal, regulatory, financial, technical, or other professional advice, nor does it constitute or override YASHADA's official AI Use Policy or any applicable government notifications, instructions, or approvals, which shall prevail in case of any duality or conflict.
References to specific AI tools, platforms, products, or services are for educational and illustrative purposes only and absolutely do not constitute endorsement, recommendation, or certification by YASHADA or its Centre of Excellence in Artificial Intelligence. Names, logos, trademarks, and screenshots remain the property of their respective owners and are used only for identification and educational purposes.
YASHADA and its Centre of Excellence in Artificial Intelligence assume no responsibility for decisions, actions, or consequences arising from the use of AI tools or reliance on AI-generated outputs referenced in this handbook.
Version 1.0 | Information reviewed as of 19 September 2026
Table of Contents
Good governance has always been shaped by its ability to discern and absorb new thoughts and opportunities into set processes of service. Today, that opportunity is Artificial Intelligence, and YASHADA, as Maharashtra's apex institute for administrative training, has, as always, taken it upon itself to ensure that our officers are not mere spectators to this transformation but its informed, capable, and responsible stakeholders.
Over the past six months, YASHADA has conducted a sustained series of training sessions on AI for officers across tiers and cadres as part of a deliberate, continuing effort to build familiarity, confidence, and judgment around AI. This initiative reflects our institutional conviction that capacity building in emerging technologies is no longer optional but foundational to good governance.
Our effort was sharpened further by the findings of a comprehensive Training Needs Analysis for Generative AI. The findings were unambiguous: there remains an enormous and largely unmet need for foundational understanding of Gen AI, particularly its everyday application in the workplace. Officers are curious, even eager, but often lack the grounding needed to use these tools with clarity and confidence. This handbook is a direct response to that gap.
It is in this context that YASHADA's Centre of Excellence (CoE) for AI, an evolving institutional capability, has been conceived. The CoE seeks to build capacity at scale, create knowledge resources that officers can turn to, and collaborate with reputed academic and technical institutions to keep our understanding current and credible. Equally important is its aim to nurture a community of practice— a network of officers who learn from each other's experiments, share what works, and hold each other to standards of ethical mindfulness and responsibility as AI becomes more deeply woven into public administration.
This handbook is one expression of that larger commitment. Rather than treating AI as an abstract or technical subject, it explains concepts through the everyday experiences of officers navigating real administrative situations, making the unfamiliar approachable, and the technical, usable.
It asks officers to become discerning users, capable of applying AI thoughtfully within the ethical and procedural framework that public service demands.
I am confident that this handbook will serve as a dependable companion for officers at every stage of their engagement with AI, and I recommend it to every officer who wishes to serve with greater efficiency, insight, responsibility, and integrity in the age of Artificial Intelligence.
Dr. Suhas Diwase, IAS
Director General
Yashwantrao Chavan Academy for Development Administration (YASHADA)
Pune
According to Greek mythology, Prometheus was a Titan who defied the Gods and took fire from Mount Olympus to give it to humanity. Fire was a divine power reserved only for the Gods, and by placing it in human hands, Prometheus transformed, forever, the destiny of the human civilization, giving us incredible power that we had to learn to use wisely because it was capable of exemplary good and tremendous harm.
This is only a legend, but today, we stand at another Promethean moment. Artificial Intelligence or AI holds a similar promise of immense transformative benefits for humanity as also the potential to wreak unimaginable destruction.
Like every radical disruption before it, it presents us with a huge challenge: unlocking its immense potential while preventing unintended consequences. That is why any capacity building in and around AI cannot be limited to technology alone but must include guardrails built into it by design.
That is also why AI literacy, which must precede its diffusion, assumes great importance and so does YASHADA’s role.
By building AI literacy and by strengthening the competency to use AI and set up robust guardrails around it, we can help ensure that public servants are mindful of not just the effective use of AI but its responsible stewardship as well, no matter what form it takes and how it develops in the future. That, precisely, is what this handbook attempts to do.
In Lewis Carroll's immortal book ‘Through the Looking-Glass,’ Alice runs alongside the Red Queen but, when she catches her breath and looks around, she realizes that she hasn’t covered any distance at all and is still where she began. Little did Lewis Carroll know that 150 years later, the Red Queen’s reply would be a metaphor for Artificial Intelligence. "It takes all the running you can do, to keep in the same place. If you want to get somewhere else, you must run at least twice as fast as that," she says to Alice.
For us, the Red Queen Effect demands a fundamental rethink. We may not be able to outpace the changes and challenges that Artificial Intelligence has brought (and will continue to bring) to the socio-economic and governance ecosystem, but we can certainly respond to them in a timely, circumspect, responsible manner.
In the world of AI, the most valuable competency is not its mastery but the ability to keep adapting and responding, to keep unlearning and learning, and to apply that learning conscientiously for the larger good.
The Red Queen never stops. Neither can we.
We hope you find this resource, brought to you by YASHADA’s Centre of Excellence in Artificial Intelligence, useful!
Team YASHADA, Pune
Maanushi, a young officer posted in a citizen facing department, is scrolling through the news feed on her phone, and notices that it seems to have posts and updates on Generative Artificial Intelligence of Gen AI oftener than before. She has heard the term on television news and in conversations with colleagues, but she doesn’t really know much about it. She decides to ask the one person who won’t mind answering her basic questions - Rajan, a colleague who likes to explore new technology and recently attended AI training at YASHADA, Pune.
Rajan is happy to orient her to the AI ecosystem Maanushi wants to know more about. ‘Think of it as rectangles inside rectangles,” he said, sketching on a notepad. “The biggest rectangle is Artificial Intelligence or AI. It’s the umbrella term: any machine or program built to do things that need human intelligence.”
“Inside that rectangle sits a smaller one: Machine Learning, or ML. This is where most of today’s AI actually happens. Instead of being programmed with fixed rules for every situation, an ML system is shown thousands of examples and it learns the pattern on its own: the better examples it sees, the better it gets.”
“Inside ML sits a smaller rectangle still: Deep Learning, or DL. It’s ML’s most powerful method, loosely modelled on how neurons in the brain connect. It needs far more data and computing power than ordinary ML, but in return it makes AI capabilities ‘impressive.’ The face-matching at Digi Yatra counters, real-time bilingual translations are all Deep Learning at work.”
“Do I not need to be tech savvy and know a bit of coding and programming to use AI?” asked Maanushi.
Rajan smiled, “Natural Language Processing, or NLP is the part that changes things for you directly: you no longer need to learn a programming language to give the computer instructions. Because of NLP, you can simply type what you want in plain English, or even Marathi, and the system understands it just the way a person would. Earlier, a programmer had to write separate code for each of these tasks. Today's NLP is almost entirely powered by Deep Learning, which is exactly why it got good enough to just talk to.”
“Wow! Unbelievable! How about Generative AI? Where does that fit?” Maanushi asked.
“That’s the newest, most talked-about slice,’ said Rajan. ‘Until recently, most AI could only recognise or predict. It couldn’t create anything new. Generative AI changed that. Built using Deep Learning, it generates original content that didn’t exist before: a draft reply, a summary paragraph, even an image or a voice clip.”
“And the AI platform we’re being asked to try — is that Generative AI?’” asked Maanushi.
“Yes,” said her colleague, “that’s an LLM, a Large Language Model. It’s a specific kind of Generative AI trained to work with language: it has read an enormous amount of text and learned what word is most likely to come next, given everything before it. So, it can draft a grievance reply, summarise a policy note, or answer a question in plain Marathi — all by predicting, one word at a time, what a good answer would look like. It doesn’t ‘know’ things the way you do, but it recognises patterns at an enormous scale.”
Maanushi looked at the notepad again: AI, holding ML, holding DL with NLP, and Generative AI, built mostly on LLMs, as the newest and most visible layer and the one she was most likely to work with.
“So,” she says slowly, “Gen AI really means a Deep Learning system, trained on language, that can draft and generate output for me.”
“That’s right,” said Rajan, “Tools like ChatGPT, Gemini, and Claude which are the three AI assistants we might use oftener than the others, ‘generate’ output as you want it, in the language of your choice, and in the form in which you need it - text, images, spreadsheets, slides, videos, audios - whatever you ask for.
"I've also been hearing the term 'agentic AI' a lot lately," Maanushi said. "Is that just another name for the same thing?"
"Not quite," Rajan replied. "Gen AI reactively creates something when you ask for it. Agentic AI goes some steps further in its scope. It can proactively and autonomously fulfil a goal. It can break it into steps, and carry out those steps on its own, researching, deciding, using other tools until the task is complete. Here, I just emailed you a digital handbook on Gen AI that YASHADA gave us after the training; I think you’ll find it useful. It's got the basics about Gen AI assistants and how to prompt them effectively. There's a whole section on step-by-step playbooks that make their application easier. Then, there’s a good chunk on what to watch out for, so you don't end up trusting something just because it sounds confident."
“Thanks,” said Maanushi, “I am sure it’ll be a great reference and handholding resource when I'm not sure how to approach Gen AI platforms and assistants.”
……………………………………………………………………………………………………………………………………………………….
And that, dear readers, is what this handbook is about! Not just helping you to use AI assistants but knowing when to reach for them and when to trust your own judgement instead.
We hope you find this resource material from YASHADA’s ‘Centre for Excellence in AI’ useful.
Part I — Safe and Responsible Use of AI Tools
Aditi is a government officer. She has been asked to prepare a short briefing note for a senior officer in her department. She is aware of and has used an AI tool that can help her with the task.
She asks the tool to draft the note. Within seconds, a polished draft appears on her screen. It is well structured. The language is good. It has even included dates, references and what appears to be a relevant Government Resolution. It looks ready. Is it?
The first question: Who is actually in charge?
Aditi could copy the draft into her document, make a few formatting changes and send it onward. But that would mean allowing the AI tool to do more than assist her. It would effectively be allowing the tool to decide what goes forward under her name.
This is where the principle of “human-in-the-loop” or HITL becomes important. Human-in-the-loop (HITL) means a human reviews, approves, or intervenes in an AI system's decision or output before it takes effect, rather than letting the AI act fully autonomously.
An AI tool can draft, summarise, organise information and suggest options. It cannot take responsibility for any of them. It has no decision-making authority and carries no accountability for what eventually leaves a government office.
Suppose the task were more consequential than Aditi's briefing note — perhaps a transfer order, a disciplinary note or a response to a citizen. An AI-generated draft becomes an official document only when a specific, named officer has reviewed it and put their own judgement and approval behind it. The moment an AI-generated draft is used without that active review, the human has quietly stepped out of the loop. A useful rule is therefore very simple:
If your name or signature goes on it, read every word first — just as you would with a draft handed to you by a junior colleague. The tool drafts. You decide.
The second question: Is everything in the draft actually true?
As Aditi reads the draft carefully, she comes across a Government Resolution number and date supplied by the AI tool. They look convincing. But Aditi does not remember seeing that particular GR in her source material. Should she trust it?
No, not until she verifies it. AI-generated answers can contain hallucinations: information that sounds plausible and authoritative but is incorrect or entirely fabricated. AI tools generate responses by predicting likely patterns in language. They do not inherently distinguish between “I know this to be true” and “this sounds like a likely answer.” That is why an invented GR number, a wrong scheme date, a non-existent judgment or a fabricated section of an Act can look just as convincing as a correct one.
So, Aditi checks the GR against the authoritative source before retaining it. The same principle applies to every specific factual claim that matters — names, dates, figures, quotations, legal provisions, policy references and citations.
Confident does not mean correct. Verify before you rely blindly.
The third question: What am I putting into the AI tool?
Aditi now wants the tool to improve one section of her note. She remembers a citizen's application that contains a useful real-life example. She is about to paste it into the AI tool. Then she stops.
The application contains the citizen's name, address and other personal information. Does the AI tool actually need any of that information to help her? This is an important part of responsible AI use: we need to think not only about what comes out of an AI tool, but also about what we put into it.
A real citizen's or colleague's story, data or circumstances should not simply be entered into an external AI tool because doing so makes a task easier. Personal information — names, Aadhaar or other ID numbers, contact details, addresses, medical information, service records and similar data — should be removed or replaced with placeholders unless the officer is working within an approved, secured government platform where such use is permitted.
If a real example is genuinely useful, anonymise it or create a fictional composite instead. Some information should not be entered at all. Official material that is confidential, restricted, classified, or covered by the Official Secrets Act must remain outside public or unapproved AI tools.
Always remove identifying personal information and use only the information necessary for your task.
The fourth question: Is the output fair?
Aditi asks the AI tool to suggest examples for her note. The examples sound reasonable, but she notices a pattern: particular occupations are repeatedly associated with men, while certain roles are associated with women. The tool has not announced that it is being biased; it has simply appeared in the output.
AI tools are trained on large volumes of existing material. As a result, they can absorb and reproduce biases that already exist in that material — including biases relating to gender, region, ethnicity etc. The AI tool may not flag this. The reviewing officer has to notice it. Responsible use therefore means asking not only, “Is this accurate?” but also “Is this fair? Whose perspective might be missing?”
Responsible AI use is therefore also about fairness and inclusion — not merely technical accuracy.
The fifth question: Are there legal implications one must be mindful about?
Aditi asks the tool a legal question relevant to her note. Can she rely on the response? She can use it as a starting point for further reading, but not as a substitute for authoritative legal advice where the issue has genuine legal significance.
Legal responsibility remains with the officer whose name or signature appears on it — not with the AI tool and not with the company that created the tool. Nor does AI create an exemption from existing law.
Copyright still applies. Passing someone else's report through an AI tool and asking it to paraphrase the text does not automatically make the material yours to publish. Similarly, the Digital Personal Data Protection (DPDP) Act, the Right to Information (RTI) Act and the Official Secrets Act continue to apply to AI-assisted work just as they apply to work produced without AI.
AI does not replace the approvals, checks or signatures that an official process requires.
The final question: Am I willing to put my name on this?
Aditi has now reviewed the draft, corrected an inaccurate reference, protected personal information, checked for bias and verified the important legal and factual claims. Her note is ready to leave her desk.
There is one last test. Would she be comfortable being personally accountable for this exact document if somebody questioned it tomorrow? This matters because an error in an AI-assisted communication, report, order, presentation or citizen response does not remain an error between an officer and an AI tool. Once it leaves the desk, it becomes the organisation's output.
A fabricated fact, inappropriate statement, biased recommendation or disclosure of information can damage not only the individual officer's credibility but also the reputation of the institution. Every officer using AI is contributing, in a small way, to how trustworthy AI-assisted government work is perceived to be.
That is why the extra minute spent checking an AI-generated output before it leaves your desk matters.
When AI Can Act, Not Just Answer
So far, we have mostly thought of AI as an assistant: we ask, it responds. It may draft a note, summarise a document, create an image or help us research a topic. But AI agents are increasingly able to do something more: take a goal, work through several steps, use other tools or sources, and take actions on our behalf.
For example, an AI assistant might be asked: “Draft an agenda for this meeting.” An AI agent might be asked to check relevant information, suggest possible dates, prepare the agenda and create a meeting invitation. The important difference is simple: the AI is moving from producing an output to autonomously carrying out a task or accomplishing a goal.
That means three questions become especially important:
1. What can it access? An AI tool may be connected to files, folders, email, calendars or other applications. Before connecting anything, consider what information you are allowing it to see. The same care you use before uploading personal, confidential or restricted information applies here too.
2. What can it do or change? Reading a document is different from editing one. Drafting an email is different from sending it. Preparing a recommendation is different from acting on it. The more consequential the action, the more important human oversight becomes.
3. What must I approve? Keep a human decision point before consequential actions. AI can prepare, organise, research and suggest. The officer still decides what is accepted, sent, changed or acted upon.
Two newer risks are worth knowing about.
When AI reads external material: Webpages, emails and documents may themselves contain incorrect, misleading or even malicious instructions. Do not assume that every instruction an AI encounters is trustworthy. External material is information to examine — not authority to obey.
When AI creates realistic images, voices or videos as deep fakes: Take care that the result does not falsely suggest that a real person said, did or participated in something that did not happen. For official or public-facing use, check that synthetic content is not misleading, identify it as AI-generated or altered where appropriate, and follow the usual approval process.
A useful rule: The more authority AI is given to act, the stronger the human oversight should be. |
Practical Guidelines for Responsible AI
Do's
1.Do treat every AI output as a draft. Read it fully before using it. Check facts, figures, names, dates and any legal or policy references against the original or another authoritative source.
2.Do protect personal information. Remove or replace names, ID numbers, contact details, addresses and other personal information with placeholders before entering material into an AI tool, unless you are using an approved, secured government platform for which such use is permitted.
3.Do keep confidential, restricted and classified material out of public or unapproved AI tools. If a document contains information protected under the Official Secrets Act or is otherwise restricted, do not paste or upload it.
4.Do retain an appropriate record of important AI-assisted work. Save your important prompts and outputs where necessary so that you can maintain the required audit trail.
5.Do disclose AI assistance where required or expected. This may be particularly relevant for training material, publications and public-facing content, depending on your office's practice and the context in which the material is being used.
6.Do keep human judgement as the final authority. AI can generate, summarise and suggest. The officer remains responsible for deciding what is used.
Don'ts
1.Do not use AI-generated content as the final version of an official communication, order, report or notice without full human review.
2.Do not enter personal data belonging to citizens, colleagues or officers into public or unapproved AI tools. This includes names, Aadhaar or other ID numbers, addresses, medical information, service records and similar personal information. Handle personal data in accordance with applicable data-protection requirements, including the DPDP framework.
3.Do not rely solely on an AI tool for legal interpretation, financial figures or policy positions. Independently verify information where accuracy or authority matters.
4.Do not assume that the AI tool's information is current. Check authoritative and up-to-date sources whenever the answer depends on a recent circular, policy change, rule, order, appointment, figure or other time-sensitive information.
5.Do not allow AI to bypass established office processes. Required approvals, checks, records and signatures remain necessary whether or not AI was used.
Remember: HITL means that the tool assists, but the human verifies, decides and remains accountable.
Part II — Prompting Fundamentals- Writing a Good Prompt:
A prompt is the instruction you give an AI tool. The quality of the response depends heavily on the clarity of that instruction.
One simple way to build a useful prompt is the CRAFT framework.
Amir has been asked to prepare a feedback questionnaire for officers who attended a two-day AI training programme. He opens an AI tool and types: “Create some feedback questions.”
The tool produces a questionnaire. It is not completely wrong, but it is not quite what Amir needs either. Some questions are too detailed, several are open-ended, and the language is not well suited to his audience. He simply needs to give the AI enough information to understand the task properly.
C — Context: First, Amir tells the AI what the situation is. “I am preparing feedback questions for officers who attended a two-day AI training programme.” Context gives the AI the background it needs: the situation, the audience, relevant facts, or documents it should take into account.
R — Role: Next, Amir tells the AI what perspective to take. “Act as an experienced training evaluator.”
A role can help guide the kind of response required from a best-in-the-field perspective,
A — Action: Now Amir says exactly what he wants the AI to do. “Create a 10-question feedback questionnaire.”
The action should be clear: draft, summarise, translate, create, analyse or whatever the task requires.
F — Format: Amir also specifies how the result should be presented. “Use 8 multiple-choice questions and 2 open-ended questions, numbered clearly.” Format might specify a table, bullet points, a word limit, or a particular number of slides, questions or sections.
T — Tone: Finally, Amir describes how the language should sound. “Keep the language simple and neutral.” Depending on the task, the tone could be formal, conversational, simple or official.
Put together, Amir's prompt now reads:
[Context] I am preparing feedback questions for officers who attended a two-day AI training.
[Role] Act as a training evaluator.
[Action] Create a 10-question feedback questionnaire that covers pedagogy, content, logistics, ambience.
[Format] Use 8 multiple-choice questions and 2 open-ended questions, numbered clearly.
[Tone] Keep the language simple and neutral.
CRAFT is not a rigid formula. The five elements do not always have to appear in this order, and every prompt will not need the same amount of detail. It is simply a useful reminder to ask:
Structured Step-by-Step Prompting
Amir's next task is different. He has prepared a draft training plan and wants AI to check whether it meets a set of prescribed requirements. Here, a simple answer such as “Yes, the plan looks fine” is not useful. Amir needs to know which requirements have been checked, which have been met, and what may be missing.
So instead of asking: “Is this training plan complete?” he tells the AI exactly what to check and in what order. This is structured step-by-step prompting. It is useful when a task involves several rules, criteria or checks because the important stages can be made visible instead of being hidden inside a single final answer.
Amir writes, ”Review the attached draft training plan against the requirements below. Report your findings step by step before giving an overall conclusion.
1.Step 1: Check whether clear learning objectives are stated.
2.Step 2: Check whether the proposed activities align with those learning objectives.
3.Step 3: Check whether the time allocated to each activity is specified.
4.Step 4: Check whether a method of assessing learning is included.
5.Step 5: Identify any requirement that is missing or unclear.
6.Step 6: Summarise the changes needed before the plan is finalised.
Use only the information in the draft and the requirements provided. Do not assume missing information.”
This approach can also be used for tasks such as checking whether a draft report contains all required sections, comparing a proposal against prescribed criteria, or reviewing whether a document follows a given checklist.
The officer should still verify the AI's findings against the original document and requirements before relying on them.
A few days later, Amir has another task. He has to draft a letter from one government department to another. This time, explaining the required style in words is difficult because the office already follows a fairly established format.
Fortunately, Amir has two good examples of similar letters. Instead of trying to describe every convention, he shows the AI the examples first. This is called few-shot prompting which means giving the AI two or three examples of the kind of pattern and output you want before asking it to produce a new one.
For Amir, this is useful because:
•It removes some of the guesswork. The AI can see the pattern rather than relying only on an instruction such as “write formally.”
•It works well for fixed formats. Government letters, orders and notices often follow conventions that can be easier to demonstrate than describe.
•It helps communicate tone. Examples can show what formal, simple, or official language looks like.
Amir has two anonymised sample letters, containing no personal, confidential or restricted information, that demonstrate the format his office follows.
Amir structures his prompt like this: “You are drafting reminder letters between government departments. I will show you two examples of a REQUEST and the LETTER written for it. Study the pattern, then write a new letter for the situation given at the end.
EXAMPLE 1 — Request: [fictional/anonymised department, subject, context] → Letter:
[anonymised sample]
EXAMPLE 2 — Request: [fictional/anonymised situation] → Letter: [anonymised sample]
Now draft the letter for this situation: [xyzzy department, subject]”
Why give two examples rather than one? With only one example, the AI may follow details too literally—for instance, carrying over the same names, numbers or wording. Two examples, anonymized as a best practice from the point of view of responsibility, make it easier to see what stays consistent and what should change. Amir is not asking the AI to copy a letter. He is showing it the pattern he wants it to follow.
Later, Amir is asked to design a half-day training module on a topic he has not worked with before.
This time, his problem is different. He knows the broad goal, but he is not yet sure about everything the AI will need to know: the participants' level, learning objectives, available time, delivery format and other requirements. He could try to anticipate every question and write one very long prompt. Instead, he asks the AI to question him first. This is flipped interaction prompting.
Normally, Amir asks the questions and the AI responds. Here, the direction is temporarily reversed: Amir states the goal, and the AI asks for the information it needs before producing the output.
He begins with: “I want to design a half-day training module on [topic] for [audience]. Before you begin, ask me one question at a time about anything you need to know to design it well (objectives, audience level, time available and format). Ask me up to five questions, then propose a module outline.”
This approach is particularly useful when:
•Amir does not yet know all the requirements himself. The AI's questions may reveal gaps he had not considered.
•He wants to avoid a wrong-direction first draft. The AI checks important details before producing a complete answer.
•The task has several moving parts or is unfamiliar. Examples include designing a new training module, planning a session or structuring an unfamiliar report.
Amir asks for one question at a time rather than many questions together. That way, each answer influences what the AI asks next. He also adds a constraint about when to stop questioning and begin drafting: “Ask me up to five questions, then propose a draft.”
Four Strategies Amir Uses to Prompt AI:
1.CRAFT helps him make an ordinary prompt complete and specific.
2.Few-shot prompting lets him use AI examples when demonstration works better than explanation.
3.Structured step-by-step prompting lets him specify the rules, checks or stages for review.
4.Flipped interaction prompting lets the AI ask questions first when Amir does not yet know all the information needed for a good answer.
When you use an AI tool, the text it works with is processed in small units called tokens. Tokens are small pieces of text that AI processes when it reads your instructions and context and generates a response.
Your instructions, relevant conversation and uploads provided to the model can contribute to the input it processes, while the response you expect contributes to the output. Your input and output, both contribute to the tokens being used up, so it helps to use tokens wisely to get more done. Let us see what this looks like in practice.
Meera Prepares a Briefing Note:
Meera needs to prepare a short briefing note for a meeting. She decides to use AI to help her.
There are several ways she could approach the task.
1. Start a Fresh Chat When the Earlier Conversation Is No Longer Relevant
Meera already has an AI conversation open from yesterday. It contains a long discussion about a completely different task. Rather than continuing there, she starts a fresh chat for today's briefing.
Long conversations can accumulate substantial previous context. AI tools manage that context in different ways, so starting a fresh chat does not translate into a predictable number of tokens saved. But if earlier material has no relevance to the new task, there is usually little reason to carry it forward.
What if Meera does need some of yesterday's discussion? She can ask AI to summarise the essential facts, decisions or instructions and use that shorter summary to begin the new task.
Keep the context you need. Leave behind what you don't.
2. Give AI the Relevant Material — Not Everything You Have:
Meera has five reports related to her subject, but the information she needs is contained mainly in one report, particularly pages 12–18. She could upload all five and ask AI to find what matters.
Instead, she provides the relevant report and points AI to the relevant section: “Use pages 12–18, particularly the section on implementation responsibilities.”
The principle applies equally to reports, spreadsheets, policies and other large files.
Providing more material does not automatically produce a better answer. Relevant material helps the AI focus on the task; unnecessary material can add context that it does not need.
Give AI enough information to do the job accurately — not everything you happen to have.
3. Be Clear the First Time:
Meera now needs three talking points from the report. She could type: “Give me some points from this report.” That is a short prompt, but the AI has to guess what she means. How many points? For what audience? How detailed? What format?
Instead, she writes: “Give me three talking points for a five-minute briefing to senior government officers. Use plain language and keep each point to no more than two sentences.”
The second prompt uses more words, but it is much more likely to produce a useful response the first time. A clear prompt that states the task, essential context and required output can be more efficient than a vague prompt followed by several rounds of clarification and correction.
Saving tokens does not mean writing the shortest possible prompt.
4. Ask for Only as Much Output as You Need
Meera needs three talking points. She does not need a detailed report, ten examples and five alternative versions. One of the most straightforward ways to avoid unnecessary output is to specify what you need:
“Give me three points on ….”
“Summarise this (upload) in 150 words.”
Ask for lengthy explanations, multiple examples or several versions when they serve a purpose and not simply because AI can generate them.
Ask for the amount of output the task actually requires.
5. Change Only What Needs Changing
Meera now has a good briefing note, but the third paragraph is too technical. She could ask: “Rewrite the briefing note.”
Instead, she says: “Keep everything else unchanged. Rewrite only paragraph 3 in simpler language.”
This avoids unnecessarily regenerating material she already intends to keep.
The same principle applies to reports, tables, emails, presentations and other outputs. If one section, paragraph, row or slide needs changing, ask AI to focus on that part.
The AI may still need access to the surrounding context to make the change correctly, so this does not mean it processes only the words being rewritten. But it does avoid generating a completely new output when only a small part needs revision.
Revise what needs revising rather than automatically regenerating everything.
6. Match the AI Feature to the Task
Meera's final task is simply to shorten a paragraph.
Her AI tool may offer advanced features such as deep research, extended reasoning or agentic capabilities. But she does not need them for a straightforward rewrite.
More advanced features are valuable when a task genuinely requires substantial research, complex reasoning, multiple steps or actions using other tools. They may involve additional model processing and searches. For simple drafting, summarising, rewriting or formatting, a standard mode may be sufficient.
Use the level of capability that the task actually requires.
Many AI tools allow users to save recurring instructions or create reusable spaces, assistants or configurations for particular kinds of work. Depending on the tool, these can retain instructions, preferences, reference material or another useful context. Features such as Projects, Gems and other reusable instruction features can support this in different ways.
The four tools in this part let you configure an assistant once — give it a name, instructions, and reference files — and then reuse it without repeating yourself every time. A free account is sufficient for everything described here; some advanced features may require a paid plan.
1. Reusable AI Assistants in ChatGPT
Set up AI once for a task you perform repeatedly
If you regularly use AI for the same kind of work, you need not explain your requirements from scratch every time. ChatGPT can be configured with reusable instructions, reference material and tools so that it approaches a recurring type of work consistently.
Example: An “Office Note Drafter” could be set up to draft official communications in formal language, follow a defined structure, and use approved templates as reference material.
Do note that Custom GPTs are currently in transition. A Custom GPT can contain:
•instructions about what it should do and how it should respond;
•a preferred tone, language and output format;
•uploaded guidelines, FAQs, manuals or templates as knowledge; and
•selected capabilities and tools, where available.
However, Custom GPTs are being phased out by OpenAI. New Custom GPTs can no longer be created on personal ChatGPT Free, Go, Plus or Pro accounts. Existing GPTs remain usable during the transition, while creation and editing continue in eligible Business, Enterprise and Edu workspaces subject to organisational permissions.
OpenAI is moving reusable workflows from Custom GPTs towards Plugins, intended to bring together reusable instructions and, where required, the apps, information and tools needed to carry out a task.
The important idea is not the name of the feature. It is to identify work that you perform repeatedly and configure AI to support it consistently. For example, you could create a reusable workflow for:
Drafting office notes | Summarising reports | Preparing training material | Reviewing documents | Creating standard communications
Instead of repeatedly telling the AI: “Use formal language, keep it concise, follow this structure, use this template…”, you configure these requirements once and reuse them.
Four things you could give your AI assistant:
•Task — What should it help you do?
•Instructions — What rules, tone, language and format should it follow?
•Knowledge — What approved templates, guidelines or reference documents should it use?
•Output — What should the finished response look like?
Then test it with realistic examples → check the output → refine the instructions → test again.
Responsible use
A configured AI assistant is still an AI assistant, not an authority. Its outputs can be incomplete or incorrect even when its instructions and reference material are well designed.
Always apply human-in-the-loop (HITL) review before using outputs for official work. Do not provide personal, confidential, sensitive, classified or restricted government information unless its use is specifically permitted under applicable government or organisational policies and within an approved AI environment.
Remember: The technology and feature names may change. The useful practice remains the same — configure AI around a well-defined recurring task, give it appropriate instructions and approved knowledge, and keep a human responsible for the final output.
A dedicated folder for one ongoing piece of work. Normally, each ChatGPT conversation is separate — nothing carries over from yesterday's chat. A Project solves this: you write standing instructions once and upload relevant files, and every conversation inside that project automatically follows them.
Example: a Project called "Training Programme 2025" containing all related files, with the instruction "All responses should be suitable for a training module aimed at entry-level government officers."
●Keeps all conversations about one piece of work in one place.
●Applies the same standing instructions to every conversation automatically.
●Stores relevant files so the AI can refer to them in any conversation inside the project.
●Lets you switch cleanly between different projects and topics.
Part A — Creating a new Project
1.Log in to chatgpt.com.
2.In the sidebar, find "Projects" and click the "+" or "New project".
3.Type a clear, descriptive name — e.g. "Budget Review 2025" — and click "Create".
1.Open the project and find the settings/gear icon or "Project instructions".
2.Type standing instructions in plain English, e.g.: "This project is for reviewing and responding to staff grievances. Always maintain a neutral, empathetic, and professional tone. Responses must be concise and clearly structured. Avoid legal language unless specifically asked."
3.Click Save.
1.Look for "Add files" or the paperclip/upload icon in the project workspace.
2.Upload background documents, data sheets, templates, or past reports.
3.These files remain available to every conversation inside the project.
Part D — Working inside your Project
1.Click "New chat" inside the project.
2.Your project instructions apply silently in the background — you will not see them on screen.
3.Type your message; ChatGPT responds according to the project's instructions and files.
4.All conversations stay saved inside the project for later reference.
Part E — Managing your Project
1.Return to project settings any time to edit instructions.
2.Add more files whenever needed using the same upload option.
3.Keep unrelated work in separate projects to avoid confusing the assistant.
Google Gemini's version of a personalised AI assistant. You create a Gem by giving it a name, instructions, and any files it needs. Once saved, it is always available in your Gemini account.
Example: a Gem called "Meeting Minutes Writer", instructed to always format notes with an agenda, discussion points, decisions taken, and action items with responsible persons and deadlines.
●Serves as a specialised assistant for any repeated task — writing, summarising, reviewing, translating.
●Follows your formatting and style preferences automatically, without re-explaining.
●Uses documents you provide as a knowledge base.
●Works seamlessly with Google Workspace (Docs, Gmail, Sheets) if you have a Google account.
Part A — Opening the Gem Manager
1.Go to gemini.google.com and sign in with a Google account.
2.In the sidebar, click "Gems".
3.Click "+ New gem" / "Create a gem".
Part B — Naming and instructing your Gem
1.Give it a descriptive name — e.g. "Drafter".
2.In the Instructions box, write clear, specific rules, e.g.: "You are an assistant that helps draft government communication. Always use formal language. Begin with 'Subject:' followed by the topic. Use numbered paragraphs. End with 'By Order' and a signature block."
3.Specify format expectations too, e.g. "Always respond in bullet points" or "Limit answers to 200 words unless I ask for more."
Part C — Uploading files (optional)
1.Use the paperclip/"Add files"/"Upload" option.
2.Add templates, reference guides, policy documents, or past examples of the expected output.
Part D — Saving and testing
1.Click Save. You will land on the Gem's conversation screen.
2.Send a test message to check the response.
3.If needed, click the pencil/Edit icon to adjust instructions, then test again.
Part E — Using it going forward
1.Each time, click "Gems" in the sidebar and select the one you need.
2.Type your request — it follows all saved instructions automatically.
3.Create as many Gems as needed, each tailored to a different task.
A Claude Project keeps standing instructions and reference files together for one piece of ongoing work, the same way ChatGPT Projects do. Best for document-heavy work where Claude needs to read and refer to your files consistently.
Example: a Project called "RTI Response Drafts" with the instruction: "You are drafting responses to Right to Information (RTI) requests. Always be factual, concise, and neutral. Structure each response with: (1) Acknowledgment of the request, (2) Relevant information or reason for denial, (3) Information about the appellate authority. Do not include information that is exempt under the RTI Act."
●Maintains a consistent working style and format across all conversations in that project.
●It lets Claude refer to your uploaded documents — policies, templates, data sheets — without you pasting them repeatedly.
●Keeps different types of work organised in separate projects.
●Saves significant time on any task you do repeatedly, week after week.
Part A — Creating a new Project
1.Go to claude.ai and log in.
2.In the sidebar, find "Projects" and click "New project" (or the "+" icon).
3.Type a clear name — e.g. "Monthly Report Summaries" — and click "Create project".
Part B — Writing your Project instructions
1.Open the project and click into "Project instructions".
2.Write standing instructions in plain, specific English.
3.Click Save, or click outside the box (autosaves in some versions).
1.Look for "Add content"/"Upload files" or the paperclip/"+" icon.
2.Add PDFs, Word documents, or text files relevant to the project.
3.Add more at any time by repeating the process.
Part D — Starting a conversation
1.Click "New conversation" inside the project.
2.Your instructions work silently in the background.
3.Type your message; Claude responds using your project instructions and files.
4.Revisit past conversations any time from the project's conversation list.
1.Update instructions any time from the project's instructions section.
2.Add or remove files through the project panel.
3.Create one project per distinct type of work — this keeps responses focused and accurate.
4.Each plan, whether free or paid, has its own project cap, so check your plan.
All four tools share the same core idea: write your instructions once, and the AI follows them every time.
Tool | Best for |
Reusable Assistant in ChatGPT | A standalone assistant with a name and personality, which can even be shared with others. |
Projects (ChatGPT) | Organising ongoing work under one roof with shared context and files. |
Gems (Gemini) | Google users who want a named, custom assistant that integrates with Google Workspace. |
Projects (Claude) | Document-heavy work where you want Claude to read and refer to your files consistently. |
Start with a tool on the platform you use most often. Once comfortable, explore the others — each has its own strengths.
Part A — Research & Content Drafting
Deep research means using AI tools to explore a topic carefully, gather information, compare ideas, and produce useful outputs like reports or summaries — no technical knowledge required.
1.Decide your topic, e.g. "I want to research how AI is being used in Indian government workplaces."
2.State your purpose clearly — learning, a presentation, a report, or understanding pros and cons.
3.Ask the tool to build a research plan: "Help me create a research plan for this topic. What should I look at?" It may suggest areas like current initiatives, challenges, good practices, and future opportunities.
4.Explore one part at a time rather than everything at once — small steps produce better results.
5.Upload your own documents (PDFs, Word files, PPTs, research papers, policy documents) and say "Use the uploaded documents while answering."
6.Share website links and say "Use this website as a reference and summarise the important points."
7.Ask follow-up questions to deepen the research: "Explain that in simple language", "Go deeper into this", "Give examples", "Compare this with another country."
8.Ask for comparisons: "Compare India's approach with Singapore's approach" or "What are the strengths and weaknesses?"
9.Ask ChatGPT to organise the information — a table, bullet points, a timeline, a comparison chart.
10.Turn the research into a useful output: a report, a presentation outline, a one-page summary, or a briefing note for senior officers.
11.Check important facts: ask "Which facts should I independently verify?"
Choosing a Platform for Research
For research, it is worth knowing how it compares to the other major tools including Perplexity, a platform built specifically for search with citations.
Dimension | Perplexity | Claude | Gemini | ChatGPT |
Primary design goal | AI-powered search with cited sources | Reasoning, writing, long-context analysis | Multimodal tasks; Google ecosystem integration | Broad general-purpose assistant |
Real-time web access | Yes — core feature, always on | Yes — available, not always on by default | Yes — via Google Search integration | Yes — available on paid plans |
Source citations | Inline citations on every response | Cites when searching; none on static responses | Cites when web grounding is enabled | Cites when browsing; not always consistent |
Long document analysis | Limited — search-oriented, not doc-deep-dive | Strong — handles large documents reliably | Large context window but consistency varies | Good; smaller context than Claude |
Dimension | Perplexity | Claude | Gemini | ChatGPT |
Reasoning / complex analysis | Moderate — relies on search more than reasoning | Strong; noted for nuanced multi-step reasoning | Good on structured tasks; less consistent overall | Strong; reasoning-mode models available |
Free tier | Yes; limited free searches per day | Yes; limited usage | Yes; fairly capable free tier | Yes; limited access to top-tier models |
Note Figures such as context window size and pricing change frequently. Treat the table above as a general orientation rather than a fixed specification, and check the platform's own site for current details before quoting numbers in official material. |
2.Drafting Reports from Multiple Sources (ChatGPT)
Use ChatGPT as a research assistant to read multiple sources, synthesise information, organise findings, and draft a professional report.
●PDF, Word, PowerPoint, spreadsheet, or text documents.
●Website links from official or credible sources.
●A clear idea of the report's audience and desired output.
1.Start a new chat and choose a model that can work with files and web content.
2.State your goal clearly, e.g.: "I need a professional report on India's AI literacy efforts in workplaces for senior government decision-makers."
3.Upload relevant documents using the attachment icon.
4.Paste official website URLs directly into the chat where supported.
5.If available, attach files from connected cloud storage (Google Drive, OneDrive, Dropbox).
6.Ask ChatGPT to identify themes, initiatives, challenges, trends, and gaps across all sources.
7.Specify the report structure you want, e.g. Executive Summary, Initiatives, Challenges, Recommendations.
8.State the tone and audience, e.g.: "Use formal, concise language suitable for senior officers."
9.Iterate: ask for expansions, summaries, tables, timelines, or rewritten sections.
10.Verify facts, figures, and claims against the original sources.
11.Copy the final output into your official Word or PowerPoint template.
Practical tips Break large requests into smaller prompts. Ask for a summary before requesting a full report. Use only authentic, official, and trusted public-domain sources. Ask for tables, infographics, and comparisons to improve readability. Always review and verify AI-generated output before sharing. |
3.Building Case Studies (Claude)
A case study is a structured, evidence-based account of how a problem was solved — challenge, approach, result. It helps officers learn from real, authentic situations and apply those lessons to their own work.
Preparation checklist — before you open Claude
●Define your goal: audience, primary message, intended use (PDF, webpage, slide deck, handout), approximate length.
●Gather raw materials: project brief, meeting notes, email threads, final deliverables, data/KPI reports, feedback or survey responses, retrospective notes, before/after comparisons.
●Prepare a short facts sheet: project/institution name, context and scale, the problem, what was done, specific results (numbers, percentages, timelines), and one to three direct quotes if available.
1.Set the context: state the audience, tone, and approximate length, and ask Claude to confirm before you share materials.
2.Paste your raw inputs (or upload the documents) in a structured way — project/topic, the problem, the approach, results and metrics, timeline, quotes, additional context.
3.Request the first draft with a defined structure: Headline, Executive Summary, Background, The Challenge, Our Approach, Results and Impact, a Quote, Conclusion/Takeaway.
4.Review the draft and give specific feedback — e.g. "Make the results section more specific and data-driven", "Restructure so the results appear earlier."
5.Once satisfied, ask Claude to format it for your delivery channel — e.g. a polished, downloadable Word document with clear headings and bolded key results.
Quality review checklist before distribution
●The headline leads with a clear, quantified outcome.
●The problem is described in plain, accurate language, not generic phrasing.
●At least one specific metric is included.
●All facts and figures have been verified against source documents.
●No confidential or protected information has been included.
●Formatting, spelling, and grammar have had a final proofread.
Part B — Assessment & Quiz Creation
4. Creating Questionnaires & Quizzes (ChatGPT)
What you need before you begin
●A topic (e.g. Cyber Safety, AI, Data Privacy, Communication Skills).
●Your audience (e.g. Students, Government Officers, Teachers, Employees).
●Number of questions needed.
●Type of questions: MCQs, True/False, Feedback Questionnaire, Survey, Assessment.
●Difficulty level: Beginner, Intermediate, Advanced (or mixed).
1.Open ChatGPT (chat.openai.com or the app) and start a new chat.
2.Decide your topic, audience, number of questions, question type, and difficulty level.
3.Type a clear instruction, e.g.: "Create 10 multiple-choice questions on cyber safety for beginners. Give 4 options for each question and mark the correct answer."
4.Read the generated questions carefully — check relevance, clarity, accuracy of answers, and difficulty level. Do not use the first output blindly.
5.Ask for improvements as needed: "Make the questions easier", "Add explanations for the correct answers", "Create scenario-based questions instead."
6.For questionnaires/surveys instead of quizzes, ask directly, e.g.: "Create a 10-question feedback questionnaire for participants who attended an AI training programme. Use objective questions and one open-ended question."
7.Specify the output format: a table, a Google-Forms-friendly layout, numbered questions, a separate answer key, or a printable format.
8.Ask for different question types as needed: MCQs, True/False, Match the Following, Short Answer, Scenario-based.
9.Fact-check important content — facts, dates, definitions, technical terms, and correct answers — especially for official assessments.
10.Copy the final output into Word, Google Forms, Microsoft Forms, slides, or a printed handout.
5. Building Interactive Quizzes with Claude
Use Claude to build the quiz as a working, self-contained web page that scores itself and can be opened on any device — no separate app or software needed.
What you need before starting
●A free Claude account (claude.ai) — no paid plan required.
●A laptop, computer, or smartphone with a browser.
●No coding knowledge required.
5.1 Creating the quiz
1.Go to claude.ai and log in (or sign up for free).
2.Click "New chat" to start with a clean, empty conversation.
3.Type a clear, specific request describing the topic, number of questions, and format — see the prompt template below.
4.Claude will generate the quiz. On a computer, it typically opens in a preview panel on the right (called an "artifact"); on a phone, tap the quiz card in the chat to open the preview.
5.Try the quiz yourself: check that the questions are clear, the marked correct answers are actually correct, the explanations make sense, and the final score screen looks right.
6.Ask for changes in plain language if needed — e.g. "make question 3 a bit harder", "change the main colour to green", "add our institute's name to the title". Claude updates the quiz and refreshes the preview automatically.
5.2 Getting the quiz onto your device
1.Once you're happy with it, find the download icon or "Download" option near the top of the preview panel and click it. This saves a single .html file, typically to your Downloads folder. If you don't see a download button, simply type: "Please give me this as a downloadable HTML file."
2.Find the downloaded file and double-click (or tap) it. It should open directly in your browser and work exactly like the preview. No internet needed once downloaded. Keep this as your master copy.
3.Do not rename the file in a way that removes the ".html" ending, or your device may not know how to open it.
5.3 Sharing the Quiz with participants
An HTML file can be sent as a WhatsApp attachment, but most participants find it easier to open a web link on their phone. Two ways to turn your file into a shareable link:
Option A — Netlify Drop (recommended, free, external site)
1.On a computer, go to app.netlify.com/drop.
2.Drag your downloaded quiz file into the box on that page.
3.Within a few seconds, a live web link ending in .netlify.app appears — this is your shareable quiz link.
4.Test the link yourself by opening it in your browser.
5.Optional: create a free Netlify account to "claim" the site so the link doesn't expire after 1 hour.
Option B — Publish directly from Claude
1.Open the quiz preview (artifact) in Claude.
2.Look for a "Share" or "Publish" button near the top of the preview.
3.Click it to generate a public link. Anyone with the link can open and use the quiz — no Claude account needed on their end.
Once you have a link, copy and paste it into WhatsApp, email, or wherever you're reaching participants — it will show as a clickable preview, and participants just tap it to open the quiz in their browser, no downloads or accounts needed on their end.
Part C — Presentation & Visual Output
1.Go to gemini.google.com and log in with a Google account.
2.Click inside the prompt box at the bottom centre of the screen.
3.Type your concept, starting with an action phrase, e.g. "Create an image of…" or "Generate a photo of…", followed by a detailed description.
4.Click Send (the paper-airplane icon) or press Enter. Gemini will produce initial baseline images in the chat.
5.Review the results. To refine, type an instruction without starting over, e.g. "Change the lighting to cool blue moonlight" or "Make the background a bit more blurred", then send again.
6.Hover over the image you prefer and click the Download icon to save it to your device.
7.Optional: for micro-edits to a specific part of the image, use the three-dot menu or Edit/Modify option on the image panel.
The same overall process, with platform-specific differences, applies for image creation on other Gen AI tools.
7. Creating PowerPoint Presentations (Claude)
1.Go to claude.ai and log in (free plan is sufficient).
2.Click inside the message box and type your request in plain English — mention the topic, number of slides, target audience, and key points.
3.Sample prompt: "Please create a PowerPoint presentation outline on the topic: 'Using AI Tools to Create Training Course Content'. Include 10 slides covering: Introduction to AI in training, Benefits of AI tools, Overview of tools like ChatGPT and Claude, Step-by-step content creation process, Quality checks, and a Conclusion. Target audience: government training officers with basic computer skills."
4.Review Claude's slide-by-slide outline. Ask for additions ("Please add a slide on…"), trims ("Make the bullet points shorter"), or tone changes ("Make it more formal / simpler").
5.Once the content looks right, type: "Now please create a downloadable PowerPoint file (.pptx) from this outline using a professional blue and white design."
6.Wait 30–60 seconds while Claude builds the file, then click the Download button that appears in the chat.
7.Open the .pptx in Microsoft PowerPoint, review every slide, and make any final edits.
Illustrative process:
1. Go to claude.ai and log in (free plan is sufficient).
Figure 1 — The claude.ai login screen. A free account is enough for this task.
2. Click inside the message box and type your request in plain English — mention the topic, number of slides, target audience, and key points.
3. Sample prompt: “Create a PowerPoint presentation outline on the topic: 'Using AI Tools to Create Training Course Content'. Include 10 slides covering: Introduction to AI in training, Benefits of AI tools, Overview of tools like ChatGPT and Claude, Step-by-step content creation process, Quality checks, and a Conclusion. Target audience: government officers with basic computer skills. I need a white background with font in Calibri point 18 black for the text and Calibri 24 bold, black for the headings. If you have any questions, ask me, one at a time, before you generate the presentation. Do not make assumptions or exceed the brief given to you.”
Figure 2 — Typing the request into the Claude message box, then pressing Send.
4. Review Claude's slide-by-slide outline. Ask for additions (“Please add a slide on…”), trims (“Make the bullet points shorter”), or tone changes (“Make it more formal / simpler”).
Figure 3 — Claude's draft outline, with a follow-up message asking for a change.
5. Once the content looks right, type: “Now please create a downloadable PowerPoint file (.pptx) from this outline using a professional blue and white design.”
6. Wait 30–60 seconds while Claude builds the file, then click the Download button in the chat.
Figure 4 — Claude builds the .pptx file; a file card with a Download button appears in the chat.
7. Open the .pptx in Microsoft PowerPoint, review every slide, and make any final edits.
Figure 5 — The finished file opened in PowerPoint, ready for a final review.
8. Converting a PowerPoint to an HTML Webpage (Claude)
This turns a .pptx file into a .html webpage — the content stays the same, only the format changes. The result opens in any browser, needs no PowerPoint installation, and is easy to share by email or WhatsApp, including working offline once downloaded.
●Have your .pptx file ready and easy to find (Desktop, Documents).
●Have a Claude account (free is fine).
●Use a computer rather than a phone — file upload and download are easier.
1.Open claude.ai, log in, and start a new chat.
2.Click the paperclip/attachment icon in the message box, select your .pptx file, and upload it (or drag and drop it into the chat window). Files over a certain size (30 MB as of the date of this documentation) may not upload.
3.Type your request, e.g.: "I have uploaded a PowerPoint file. Please convert it into a well-designed HTML webpage. The HTML file should display all the slide content clearly — including headings, bullet points, and any text from each slide. Use a clean, professional design with a navy blue and white colour scheme. Make it easy to read on a computer screen. Please give me a downloadable .html file."
4.Press Enter and wait while Claude reads the file and generates the HTML.
5.Review the preview or generated code: check that all slide content is present. Note that complex animations, embedded custom fonts, and precise pixel layouts may not transfer.
6.Ask for specific changes if needed, e.g. "Slide 4 on [topic] seems to be missing. Please add it" or "Please change the colour scheme to blue and white."
7.Click the Download button in Claude's response. If none appears, ask Claude to give you the complete HTML code in one block, then paste it into Notepad and save it with a .html extension (choosing "All Files" as the save type).
8.Open the saved file in your browser to check the final result, then share it by email, WhatsApp, or cloud storage.
Part D — Live Session & Engagement Tools
9.Running Live Polls & Word Clouds (Mentimeter)
Mentimeter (menti.com) lets you run live polls, word clouds, rankings, and quizzes that participants answer on their own phones in real time, with results appearing instantly on the shared screen for icebreakers, checking understanding mid-session, or a quick anonymous pulse-check.
Getting set up
1.Go to www.menti.com in your browser.
2.Click "Sign up" and create an account using your work email and a password or log in if you already have one. Signing in with a separate email/password rather than your Google account is safer.
3.Choose the Free plan to start — it supports up to 50 participants per month, which is enough for most single-session training use.
Creating your first slide — a word cloud
1.From the home page, click "New Menti" — a panel of question types opens (Word Cloud, Multiple Choice, Open Ended, Scales, Ranking, Q&A, and more).
2.Click "Word Cloud".
3.In the input box, type your question e.g. "What word comes to mind when you think of training?"
4.Participant responses appear live on screen as a word cloud, with more frequent words shown larger.
Sharing it with participants
1.Click the share icon near the top of the editor.
2.A panel opens showing your participation link and QR code.
3.Set the access code expiration — the default is 2 days, but 7 or 14 days is a better choice if the same Menti will be reused across a multi-day programme.
4.Copy the link to share by WhatsApp or email, or download the QR code to paste onto a slide for participants to scan in the room.
Running it live:
1.Click "Preview" first to check exactly what participants will see — preview responses are not saved.
2.When ready, click "Present."
3.Use the timer control at the bottom of the slide if you want to cap response time e.g. 1–2 minutes.
Other Mentimeter slide types worth knowing:
●Ranking: participants order items (e.g. "Where can AI tools help you the most”) as a live bar chart.
●Pin on Image: participants click a point on an uploaded image or icon set (e.g. "Which AI tool do you use the most?").
●2×2 Grid: participants place items along two axes (e.g. impact vs. effort) to visualise priorities
●Open-Ended Questions: free-text responses appear live; useful for thought provoking prompts
Using Mentimeter - A Visual Guide:
10. Collaborative Boards for Engagement & Assessment (Padlet)
Padlet (padlet.com) is a shared digital board where participants post text, images, or files that everyone can see build up in real time — useful for brainstorming, exit reflections, anonymous feedback walls, or organising group work into sections.
Getting started
1.Search for or go directly to padlet.com in your browser.
2.Choose "Sign up" (new account) or "Log in" (existing account).
3.Sign in with your work email and password.
Creating your board
1.From the home page, click "+ Make" to start a new Padlet.
2.Choose a format that fits your activity. "Wall" is the simplest general-purpose choice for a running feed of posts; "Columns", "Table", "Timeline", and "Map" suit more structured activities; "Whiteboard" and "Brainstorming web" suit open visual work.
3.A default title appears — click to rename it to something specific to your session, e.g. "Facilitation Skills — My One Takeaway".
Sharing your board
1.Click "Share" near the top of the board.
2.Set the participant access level. "Writer" (visitors can post) is usually the right choice for a live activity; "Reader" only allows viewing; "Moderator" lets visitors approve posts too.
3.Copy the shareable link, or click the QR code option to get a scannable code you can paste onto a slide.
4.Decide on privacy options — in particular, whether to require visitors to log in. Leaving this off keeps responses anonymous, which is often better for honest feedback.
Optional — breakout rooms for differentiated feedback
If you need feedback from different sub-groups on different topics, add multiple sections to the same board (e.g. "Session Plans", "Assessments", "Activities") and use Padlet's breakout rooms feature to give each section its own shareable link. Each group then sees and posts to only their own section.
After the session
1.Once participants have responded, use "Export as PDF" (found in the share panel) to save the filled board as a handout or slide deck for documentation.
2.This exported PDF can then be uploaded to an AI tool (see the Reports and Deep Research playbooks earlier in this handbook) to summarise themes, sentiment, and common takeaways across all responses — turning a wall of sticky notes into a structured summary in minutes.
Reminder As with any tool that collects participant input, keep boards anonymous where appropriate, and review AI-generated summaries of participant feedback before circulating them, the same way you would review any other AI output in this handbook. |
Using Padlet: A Visual Guide:
Step 1
Getting Started: Find Padlet online, and choose to sign up or log in
1. Access Padlet on your browser
2. Sign up (if you do not have an account already) or log in (if you do)
3. Sign in with email/user name and password
Step 2
Go To Home Page — Make a Padlet
Once logged in, click “+ Make” to start a new Padlet.
Step 3
Choose a Format: Select a board or sandbox format that suits your activity.
Step 4:
Set Up Your Wall: A default title appears. You can give your Padlet a title and select the “Wall” format.
Step 5
Share Your Board: Rename your Padlet and click “Share” to set permissions and get a shareable link for participants as well as a QR code that you can paste in a PPT.
Step 6
Choose the participant access: Choose the right access for your participants. Usually, Writer is a suitable choice.
Step 7
Optional: Breakout Rooms for Differentiated Feedback from different sections of participants: You can create separate sections that can be used for breakout rooms.
Step 8:
A Filled-In Padlet Wall (sample): This is how one of my previous Padlet walls looked when it was filled with participant responses. I chose settings in which participant names were not recorded to keep the responses anonymous.
Step 9
Sharing the Filled-In Board: After all the participants have responded, you can share the completed Wall or board for review and analysis.
Step 10
Export Padlet as PDF: One way of doing that is to export the Wall as a PDF handout or slides for documentation and to feed into AI tools for analysis.
Part E — Video & Multimedia Output
11. Creating Video Overviews from Documents (Gemini Notebook)
Gemini Notebook (notebook.google.com) can turn a set of source documents into a narrated video. Gemini Notebook generates the script, narration, and visuals itself from your source documents.
Phase 1 — Preparing your sources
1.Go to notebooklm.google.com and click "+ Create New" to start a new notebook.
2.In the box that appears ("Create Audio and Video Overviews from…"), add your sources — upload documents from your saved files or Drive, add web links, or type in text directly. For best video quality, use clean Google Docs or clean PDFs rather than messy scans.
3.Alternatively, close that box and use "Add sources" (the + icon) on the left. You can select up to 50 sources. Make sure each source's checkbox is checked.
Phase 2 — Entering the AI Studio
1.On the right-hand side of the screen, click the "Studio" tab.
2.You will see several generation cards — Audio, FAQ, Study Guide, and others. Find the card titled "Video Overview".
Key step Do not just click "Generate." Click the pencil icon (or "Customize" button) at the bottom of the Video card first — this unlocks the settings that make the output usable. |
Phase 3 — Customising style, language, and quality
Clicking Customize opens four settings:
●Format — "Explainer" is best for teaching connections across multiple documents; "Brief" gives a quick 1-minute highlight reel; "Cinematic" (available to Pro and Ultra users only) produces a fully animated, story-driven video.
●Visual style — choose "Whiteboard" for a hand-drawn tutorial feel, "Anime/Watercolour" for a more artistic look, or "Auto" to let the AI decide.
●Language — English is the default; over 80 languages are available, including Marathi. This changes both the voiceover and the on-screen text.
●Audience instruction box — tell the AI exactly who the video is for. A simple formula works well: "Target: [audience]. Tone: [tone]. Goal: [what the video should achieve]."
Phase 4 — Generating and downloading
1.Click the blue "Generate" button. Gemini Notebook will build a script and generate visuals — this typically takes 3 to 10 minutes.
2.You can close the window and return later; a progress bar remains visible in the Studio panel.
3.Once finished, the video appears in the Studio panel. Click the download (arrow) icon to save the file as an MP4 to your computer.
Reminder: As with every tool in this handbook, review the generated video before sharing it — check facts, terminology, and tone — and do not upload documents containing confidential or personal data as sources.
12. Translating Documents — Marathi ↔ English
Several tools are available for translation. All machine translations must be reviewed before issuing official correspondence with attention to government terms, legal references, names, and dates.
If you need to… | Use this tool | For Marathi |
Translate inside a Word, PPT, or Excel file | MS Office (Review tab) | Good |
Translate an official communication or notice | Bhashini (Anuvaad): bhashini.gov.in | Excellent |
Quick translation of a short text or image | Google Translate | Good |
Translate and improve tone, formality, or clarity | ChatGPT / Gemini / Claude | Good |
Translate spoken speech (voice input) | Bhashini or Google Translate | Excellent |
A. Microsoft Word, PowerPoint, or Excel
●Select the text, click the Review tab, then Translate → Translate Selection (or Translate Document for the whole file).
●The Translator pane opens on the right; confirm the detected source language and choose your target language (Marathi or English).
●Review the translation, then click Insert to place it, or copy-paste it manually.
●In Excel, keep the original in one column and the translation in another rather than overwriting the source data.
B. Bhashini (Anuvaad) — Government of India language platform
●Go to bhashini.gov.in and open Anuvaad, then Text to Text.
●Set the source and target languages and type or paste your text.
●The translation appears automatically; copy it into your document.
●Use the swap icon (↔) to reverse direction (Marathi → English).
●Best for official government correspondence, circulars, notices, RTI responses — Bhashini's models are trained on government and public-sector language.
●Text: go to translate.google.com, set source/target language, and type or paste. translation appears automatically.
●Document: click "Documents", upload a Word/PDF/PPT file, select languages, and click Translate, then download the result.
●Image: click "Images" and upload a photo of text (e.g. a scanned notice) for on-the-spot translation.
●Voice: click the microphone icon and speak in Marathi or English for instant transcription and translation.
D. ChatGPT, Gemini, or Claude — translation plus tone control
Unlike other tools, AI assistants can translate and adjust the tone at the same time — more formal, simpler for citizens, or better suited to official correspondence.
Answers to the problems that come up most often when working with AI assistants:
Accounts & Access
Q: I can't remember which email I used to sign up.
A: Try the "Forgot password" link on the platform's login page with each email you might have used (personal and official). Going forward, pick one email — ideally your official email address where the platform allows it — and use it consistently across all these tools.
Q: The tool says I've reached my daily or monthly limit.
A: Free plans cap usage — for example, ChatGPT and Claude free tiers limit how many messages you can send per day, and Mentimeter's free plan caps at 50 participants a month. Wait a few hours or until the next day/month, or ask your organization’s point of contact whether an institutional paid plan is available for heavier use.
Q: Should I sign up with my personal email or my official email?
A: Use your official email where the platform accepts it — it keeps the account recoverable if you change personal email providers, and keeps departmental work separate from personal accounts. Check with your organization if you're unsure whether an institutional account already exists for a given platform.
Uploading Files
Q: My file won't upload, or the upload gets stuck.
A: Most tools cap file size — Claude's free plan, for instance, may reject files over roughly 25 MB. Try removing unnecessary images from the file first, or split a large document into smaller sections and upload them one at a time.
Q: The tool says my file format isn't supported.
A: Convert the file to a more common format first — PDF or Word (.docx) for documents, PNG or JPG for images. Most tools in this handbook accept these without issue.
Q: I uploaded a scanned document and the results are poor.
A: Scanned pages (essentially photographs of text) are harder for AI tools to read accurately than a native digital document. Where possible, use a clean Word or Google Docs file, or a PDF exported directly from a document rather than scanned from paper.
Getting Good Output
Q: The AI refused to do what I asked.
A: This usually means the request was ambiguous in a way that resembled something the tool declines, or it touched a topic the tool is cautious about. Rephrase with more specific, legitimate context — e.g. name the training purpose and audience explicitly — and try again. If it still declines, the task may fall outside what's appropriate to ask an AI tool to do (see Part I, Section 3).
Q: The output looks confident but I suspect it's wrong.
A: This is called hallucination (see Part I, Section 1) — treat it as a normal, expected limitation rather than a malfunction. Verify names, dates, figures, and claims against your own source documents, or ask the tool directly: "What is this based on?" and check that source yourself.
Q: The formatting or structure is messy.
A: Don't start over — just ask directly: "Reformat this as bullet points", "Add clear headings", or "Put this in a table." Every tool in this handbook can revise its own output through a follow-up instruction in the same conversation.
Q: I don't see a download button for the file I need.
A: Type the request directly: "Please give me this as a downloadable [Word / PowerPoint / HTML] file." This works across ChatGPT, Gemini, and Claude and is mentioned in several playbooks in Part IV.
Q: I lost my work, or the conversation disappeared.
A: Check the sidebar or chat history — most tools save conversations automatically under your account. For anything important, it's still good practice to copy key drafts into a separate document as you go, since it can be hard to locate one specific version inside a long back-and-forth conversation.
Sharing & Links
Q: My Netlify quiz link has stopped working.
A: Unclaimed Netlify Drop links can expire after 1 hour (see IV.2A). Either drag the file into app.netlify.com/drop again for a fresh link, or create a free Netlify account and "claim" the site so the link stays permanent.
Q: The QR code isn't scanning properly.
A: Check that it hasn't been compressed or shrunk too much when pasted onto a slide — keep it reasonably large and high-contrast. Always test-scan it yourself before the session, and have the plain-text link ready as a backup.
Q: My Mentimeter access code expired partway through a multi-day programme.
A: Set the access code expiration to 7 or 14 days instead of the 2-day default when you expect to reuse the same Menti across multiple sessions (see IV.10).
Q: I've hit Mentimeter's participant limit or Padlet's board limit.
A: Mentimeter's free plan resets its 50-participants-a-month count monthly; Padlet's free plan caps the number of active Padlets you can have at once. Archive or delete old boards you no longer need, plan session sizes accordingly, or check with your organization about an institutional paid plan.
On the Day of the Session
Q: There's no internet in the training hall.
A: Most tools in this handbook need internet while you're creating or running content live. Files you've already downloaded — an MP4 video, an HTML quiz, a .pptx — will open and work offline once saved, so download anything you'll need before you lose connectivity.
Q: Participants can't open the link on their phones.
A: Confirm the sharing/access setting is public or "anyone with the link" rather than restricted to specific accounts, and test the link yourself on a phone before the session, not just on your laptop.
Safety & Data
Q: I accidentally pasted something I shouldn't have — a real name, confidential figures.
A: Delete the message or conversation where the platform allows it. Treat this the same way you would treat accidentally emailing something externally — if your department's policy requires reporting a data-handling slip, do so, since the platform may retain the input even after you delete it from your own view.
If you need to… | Platform |
For letters, notes, and other writing | ChatGPT, Gemini, Claude |
Set up an assistant for a recurring task | ChatGPT, Gemini (Gems), Claude (Skills, Project) |
Turn source documents into a report | ChatGPT, Gemini, Claude |
For MCQs, quizzes, questionnaires | ChatGPT, Gemini |
Create an image or graphic | Gemini, ChatGPT, Canva |
Research a topic in depth | ChatGPT, Claude, Perplexity, Gemini |
Build a structured case study | ChatGPT, Claude, Perplexity, Gemini |
Create a PowerPoint presentation | Claude (editable), Gamma, Canva, ChatGPT |
Convert a PowerPoint into a shareable webpage | Claude |
Translate a document between Marathi and English | MS Office, Bhashini, Google Translate |
Conduct real-time polls, word clouds, quizzes | Mentimeter |
Have participants respond on a collaborative wall | Padlet, Slido, Mentimeter |
Platform | Web address |
ChatGPT | |
Gemini | |
Claude | |
Bhashini (Anuvaad) | |
Gemini Notebook | |
Mentimeter | |
Padlet | |
Slido | |
Canva |
AI Policies and Guidelines:
Appendix C: Glossary of Useful Terms
1.Artificial Intelligence (AI) — The use of computers to simulate or replicate human-like thinking and behaviour. The term encompasses a wide spectrum of applications, from sophisticated conversational systems to simple programmed behaviours in video games.
2.Machine Learning (ML) — A branch of artificial intelligence in which programs learn to make predictions or classifications by identifying patterns in large datasets. Rather than being explicitly programmed with rules, a machine learning model is "trained" on data to produce its own outputs.
3.Deep Learning — A subset of machine learning that relies on large, layered structures called neural networks, which require vast amounts of data to function. Deep learning became dominant in the field around 2012 and has continued to advance rapidly since.
4.Large Language Model (LLM) — A type of deep learning model trained to predict the next word in a sequence of text. LLMs such as GPT, Claude, and Gemini are the foundation of modern AI assistants and, while capable of a broad range of tasks, are rooted in this next-word-prediction architecture.
5.AI Model — A computer system trained on large amounts of data to recognise patterns and produce outputs. Different AI models are designed or optimised for different tasks, such as generating text, analysing images, translating languages, or reasoning through problems.
6.Foundation Model — A large AI model trained on broad, general-purpose data that can then be adapted or fine-tuned for many different downstream tasks, rather than being built to do just one thing. Large language models are one type of foundation model.
7.Frontier AI — A term referring to the most capable and recently released large language models from leading AI research organizations. Because the field advances quickly, "frontier" describes whichever models are currently at the cutting edge of performance.
8.Natural Language Processing (NLP) — The broader field of AI concerned with enabling computers to understand, interpret, and generate human language. Large language models represent the current dominant approach within this field.
9.Training Data — The information used to teach an AI model to recognise patterns and generate responses. The quality, quantity, and variety of this data can influence how well the model performs and what biases or limitations it may develop.
10.Generative AI — A category of AI that creates new content, such as text, images, audio, or code, rather than simply classifying or predicting a category from existing data. Large language models are the most widely used form of generative AI today.
11.Multimodal AI — AI that can work with more than one type of information, or “mode,” such as text, images, audio, video, or documents. A multimodal AI assistant may, for example, analyse an uploaded image as well as respond to written instructions. The handbook itself refers to multimodal capability when comparing research platforms.
12.AI Assistant — An AI-based tool designed to interact with a user and assist with tasks such as drafting, summarising, analysing, translating, researching, or creating content. Some assistants can also retain instructions or reference material for recurring kinds of work.
13.Chatbot — An AI interface that communicates with users exclusively through text. The user submits a text input (called a prompt), and the model responds in kind. All interaction is confined to the exchange of written language.
14.Prompt — The text input, instruction, or question a user gives to an AI model to elicit a response. The quality and clarity of a prompt strongly influence the quality of the model's output, which is why "prompting" is often treated as a skill in itself.
15.Token — The basic unit of text an AI model processes internally, roughly corresponding to a word or part of a word. Models read input and generate output one token at a time, and factors such as context window size and usage limits are typically measured in tokens.
16.Context Window — The amount of information an AI model can work with at one time while generating a response. A larger context window allows the model to consider more text, conversation history, or document content together.
17.Agent (Agentic AI) — An AI system that can both communicate with users and take actions in the world, such as sending emails or browsing the web. Unlike a chatbot, an agent can receive instructions, execute tasks, and report results. The adjective form, agentic, describes any AI behavior involving real-world action or task completion.
18.Guardrails — Rules, filters, or design constraints built into an AI system to limit what it will do or say. Guardrails are used to reduce harmful, inaccurate, or inappropriate outputs and to keep the system's behaviour within intended boundaries.
19.Bias (Algorithmic Bias) — A systematic skew in an AI system's outputs, often traceable to patterns or imbalances in the data it was trained on. Bias can lead to outputs that unfairly favour or disadvantage particular groups, and is a key concern in responsible AI use.
20.Hallucination — An instance where an AI model generates plausible-sounding but factually incorrect or fabricated information, often stated with the same confidence and fluency as accurate output. Recognising this tendency is essential to using AI systems responsibly.
21.Human-in-the-Loop (HITL) — An approach in which a human reviews, verifies, approves, or intervenes in an AI system's output or actions rather than allowing the AI to operate entirely on its own. In official work, this ensures that human judgement and accountability remain central.
22.Anonymisation — The process of removing or changing information that could identify a particular person before data is shared or processed. When using public AI tools, this can include replacing names, identification numbers, addresses, and other identifying details with neutral placeholders.
Three months later, Maanushi is no longer scrolling uncertainly past newsfeeds about Generative AI; she’s the one others in her workplace turn to when they need help figuring out how best to use AI capabilities.
She has integrated AI assistants in her workflow, but she uses them mindfully. She scrutinizes and verifies every word of what AI generates because she is aware that she ‘owns’ the responsibility of whatever she sends out or uses.
When a new officer in her division asks her the same questions that she’d once asked Rajan, she finds herself reaching for a notepad, happy and willing to pass on what she knows.
Maanushi has not suddenly become a technical expert or a master at using AI assistants. It is just that she does not fear failure when she experiments with them. More importantly, she has become a ‘learner,’ keeping herself abreast of updates and developments that matter to her personally and at work.
That habit is worth carrying beyond this handbook. The tools themselves will keep changing; what won’t change is the value of staying curious about them. A fifteen-minute podcast episode on your commute, a YouTube walkthrough of a feature you haven’t tried, a short article that explains one new update — these add up faster than any single training session.
Most of all, practise without worrying about getting it wrong. A clumsy first prompt, a draft that needs three rounds of editing, a tool that doesn’t work the way you expected — none of that is failure.
It’s simply what learning new technology looks like before you become conversant, comfortable, and competent in using it.
That’s the real takeaway this handbook offers: not a fixed set of instructions (because they are likely to become outdated with time), but the motivation to keep unlearning and learning all through our professional and personal lives.
……………………………………………………………………………………………………………………………………………………………..