Quick Answer
A good prompt for slides or a document describes the finished file first: who reads it, what they must decide, how many slides or pages, and the structure of each part. Then it limits the AI to your own facts, asks it to show calculations, and has it flag gaps instead of filling them.
The best way to get a usable slide deck or document from AI is to describe the finished file before you describe the topic. Say who will read it, what they need to decide, how many slides or pages it should have and what goes in each one. Then tell the model to use only the facts you give it and to flag anything missing.
Most people do the opposite. They type "make a presentation about our Q3 sales" and get ten slides of confident filler, with numbers that are close to theirs but not quite the same. Fixing that takes longer than writing it yourself, which is why plenty of people try AI for documents once and go back to a blank page.
This matters because documents are what most professionals in the region use AI for. Twenty prompts went into the library this week to make that work better. They cover all ten categories, two each, and every one produces something you can send, print or present.
What Are the 20 New Prompts For?
They fall into four groups.
Presentations. Write Speaker Notes for a Finished Slide Deck, Turn a Sales Spreadsheet Into a 5-Slide Monthly Review, Turn a Lesson Plan Into Student-Facing Slides, Prepare a Case Study Presentation for an Interview, Turn Your Year's Results Into a Performance Review Deck, and Create Matching Background Images for a Slide Deck.
Documents people sign or act on. Turn a Long Report Into a Two-Page Brief, Build a Client-Ready Quote Document From Line Items, Build a Printable Worksheet With an Answer Key, Turn Monthly Sales Into a Bank Loan Summary, Check a Supplier Contract for Personal Data Clauses, Design a Cover Image for a Report or Proposal, and Turn a Research Session Into a Shareable One-Pager.
Getting documents to people. Write a Cover Email for an Attached Deck or Report, Ask a Client for Feedback on a Draft Document, Turn One Slide Deck Into a Week of Posts, and Turn Customer Survey Answers Into Social Proof Posts.
Doing it again next week. Write a Python Script That Turns a CSV Into Charts, Check an AI-Generated Script Before Running It, and Save a Repeat Task as Reusable AI Instructions.
You'll find them in their categories, with the two most regional ones, the bank loan summary and the data clause check, in SEA Business. The quote and monthly review prompts are in Business.
Why Do Slide and Document Prompts Need a Different Approach?
Because a document has a reader and a job, and a chat answer usually doesn't.
When you ask a question in a chat, a decent answer is enough. When you ask for a deck, the output has to survive someone else reading it: a manager deciding on budget, a bank officer judging a loan, a client checking a quote line by line. That changes what "good" means. Length, order and accuracy matter more than eloquence.
It's also where most of the use already is. Robert Walters' report The Jobs of Tomorrow: How AI is Reshaping Work Across Asia surveyed close to 5,000 professionals and companies across 10 Asian markets, including Singapore, Malaysia, Indonesia, the Philippines, Thailand and Vietnam. It found 93% of professionals are actively using AI at work, and that the most common uses are content creation (47%), data analysis and insights (40%) and research (38%). Only 58% of employers had formally introduced AI, which means a lot of people are working this out on their own.
The tools have caught up, too. Anthropic says people have made more than 45 million docs, decks and designs in Claude, and its announcement says Docs, Slides and Design are now available on every plan, including Free. When the AI can hand you an actual PowerPoint file, a vague prompt produces a vague file, and that's harder to fix than a vague paragraph.
How Do You Describe the Finished File Instead of the Topic?
Answer five questions in the prompt before you ask for content:
- Who reads it? "The two founders" and "a bank credit officer" need very different documents from the same sales data.
- What must they decide or do? Approve a pilot, sign a quote, release a loan. Every slide should push toward that.
- How big is it? Five slides, two pages, under 120 words. Size limits do more for quality than any instruction about tone.
- What goes where? A slide-by-slide or section-by-section structure, with one job per slide.
- What does it look like? A4 portrait, 16:9, no more than 25 words on a slide, prints on two pages.
The monthly review prompt shows this in practice. Slide 1 is the headline result against target in one sentence and one number. Slide 2 is what drove it. Slide 3 is what dragged it down. Slide 4 is one chart, and the model has to say which chart type and why. Slide 5 is three actions with owners. There's very little room for filler, because every slide already has a job.
The student slides prompt does the same thing differently. It caps on-screen text at 25 words and moves everything else into speaker notes. That one rule fixes the most common problem with AI-made teaching slides, which is that they're lecture notes pasted onto a background.
How Do You Stop Numbers and Facts From Drifting?
Give the model three rules, and repeat them in every document prompt.
Use only what I give you. The bank loan prompt says "Use only the figures I provided." The quote prompt says to leave the tax line blank rather than guess a rate. Without that, models tend to round, estimate or borrow a plausible industry average, and a plausible number in a loan application is worse than a missing one.
Show your working. The monthly review asks for an appendix of calculations, and the bank summary asks for the margin calculation to be shown. Most spreadsheet errors in AI output come from a misread column or a wrong row range, and a calculation appendix exposes that in seconds.
Point back to the source. The two-page brief prompt asks for the page number behind every finding. The research one-pager asks the model to label any claim without a source as "Unsourced" instead of quietly confirming or dropping it. You can check a referenced document in five minutes. You can't check an unreferenced one without redoing the work.
This is the same discipline we covered for images in last week's multimodal prompts guide: make the model show you what it read before it tells you what it thinks.
When Should You Ask for the Structure Before the Content?
Whenever the document is longer than a page or a few slides, or someone important will read it.
Asking for an outline first costs one extra message and saves a full rewrite. You see the order of the argument while it's still cheap to change, and the model writes the full version against a structure you've already agreed. The interview case study prompt does this by asking for a slide-by-slide outline with one headline per slide before anything else. The headlines alone should tell the story. If they don't, the deck won't either.
This is a simple form of prompt chaining, where each step's output becomes the next step's input. If you want to go further, our prompt chaining guide covers longer chains for reports and proposals.
For short pieces, such as a cover email or a quote, skip the outline. The prompt's structure is the outline.
Which AI Tools Can Turn a Prompt Into an Actual File?
Not long ago the answer was mostly "copy and paste". Now several tools produce a file directly.
Claude's Slides and Docs can export decks to PowerPoint and PDF, or send them to Google Slides as editable files, according to Anthropic. WhichAIBest's Claude review covers what the Free plan now includes. Gamma builds a designed deck from an outline, and our Gamma prompts guide covers how to brief it. ChatGPT, Claude and Gemini can all write the Python that turns a CSV into chart images you drop onto slides, which is what the new chart script prompt is for.
Two cautions. First, these tools are only as good as the brief, so the five questions above still apply. Second, check whether your company allows client data in the tool you're using. That's especially true for the contract and loan prompts, which involve personal and financial data that Singapore's PDPA, Malaysia's PDPA, Indonesia's PDP Law and Thailand's PDPA all regulate.
For smaller firms, the first file is often the hardest step. IMDA's Singapore Digital Economy Report 2025 found AI adoption among Singapore SMEs rose from 4.2% to 14.5% in a year, mostly through off-the-shelf generative AI tools. Ready-made prompts that produce a real deliverable are an easy way to be part of that number.
How Should You Use These This Week?
Pick one document you make every week or month and run it through the matching prompt. A sales review, a client update or a quote are good candidates, because you already know what a good one looks like and you'll spot problems fast.
If the result is close, save it with the new reusable instructions prompt. It turns a task you repeat into instructions you can store as a Gemini skill, a Claude project or a ChatGPT project, so next time you only paste in the new inputs.
And before you send anything, run the numbers once yourself. The prompts are built to make that check quick. They don't make it optional.
What Else Do People Ask?
Do these prompts work in ChatGPT, Claude and Gemini?
Yes. Every prompt in this batch is model-neutral and works as plain text in ChatGPT, Claude and Gemini. The ones that start from an attachment, such as a spreadsheet, report or slide deck, need a plan that allows file uploads. If your tool can export slides or documents directly, paste the output into that feature. Otherwise, copy the text into PowerPoint, Google Slides or Word.
How many slides should I ask AI for?
Fewer than you think. Ask for the smallest number that covers what your reader must decide, often five to eight slides for an internal review or interview case. Give each slide one job and a word limit. A fixed slide count with a defined purpose for each slide does more for quality than any instruction about tone or style.
Why does AI change the numbers in my document?
Usually because the prompt let it. Without a rule, models round figures, estimate missing values or misread a spreadsheet column. Tell the model to use only the figures you provide, to show its calculations in an appendix, and to point to the page or row each number comes from. That makes errors visible in minutes instead of after you send the file.
Is it safe to upload client documents to AI tools?
It depends on your plan's data settings, your company's policy and your contract with the client, so check all three first. Contracts, payroll data and loan figures contain personal or financial data covered by laws such as Singapore's PDPA, Malaysia's PDPA, Indonesia's PDP Law and Thailand's PDPA. When in doubt, remove names and identifiers before uploading.
Should I edit the AI's draft or write a better prompt?
Fix the prompt when the structure, length or audience is wrong, because editing will not save a document built on the wrong frame. Edit the draft by hand when the structure is right and only wording or details need work. If you make the same fix twice, add it to the prompt or save it as reusable instructions so the next draft starts closer.
Sources: Robert Walters, The Jobs of Tomorrow: How AI is Reshaping Work Across Asia, for AI use at work, top use cases and employer adoption across 10 Asian markets. Anthropic, Dashboards and Motion announcement, for Docs, Slides and Design availability, exports and the 45 million figure. IMDA, Singapore Digital Economy Report 2025, for SME AI adoption. Tool features change often, so check each vendor before relying on an export option.