[90% Time Reduction] Ex-Goldman Sachs Kei Tanaka's Genius AI Slide Creation Method

People who are slow at creating materials open PowerPoint first.

They stare at a blank slide, place a title, and move shapes around while trying to think about "what to convey." This means they are trying to process thinking, structure, and design all at once. It is only natural that it takes time.

The method demonstrated by Kei Tanaka, formerly of Goldman Sachs, follows a completely different order.

1. Output thoughts via voice.
2. Build logic with Claude.
3. Convert to slides with Marp.

This is the overall picture.

In this article, based on Mr. Tanaka's demonstration and interview, I will break down the following into a usable format:

• The slide creation flow introduced as "90% time reduction"
• The true nature of Mr. Tanaka's "templates" that generate speed
• Three copy-paste prompts for creating materials from voice memos

This is not just a tool introduction. It is a work technique that rearranges the very order of creating materials.

Why Kei Tanaka's AI Slide Creation is Abnormally Fast

To put it simply, Mr. Tanaka does not create slides one by one.

He first gathers the information in his head, outputs it, and lets AI convert the structure and format.

Mr. Tanaka worked at Goldman Sachs for 17 years and served as the co-head of the investment division. In his current investment work, he says he uses AI for research, hypothesis building, decision support, and material creation.

In interviews, he explained that complex research and material creation that previously took about three weeks now takes about one hour when combined with AI.

He also revealed a method where he speaks about five A4 pages worth of ideas in about two minutes and lets AI handle the transcription and organization.

Source: The Keyperson "Kei Tanaka"

https://thekeyperson.biz/interview/tanakakei/

Furthermore, in a video by Digirise Co., Ltd., Mr. Tanaka demonstrated slide creation using Claude and Marp. The video title claimed a "90% time reduction."

Source: Chaen [AI Research Institute] "Ex-Goldman Investment Division Japan Top Demonstrates! AI Work Techniques Achieving 90% Time Reduction"

https://www.youtube.com/watch?v=C5qAt69rA54

Please do not misunderstand.

90% is not a guaranteed figure for everyone or every document. It is a demonstration that was possible because of Mr. Tanaka's experience, the decision criteria in his head, and the amount of information he could provide to the AI.

So, is it impossible for ordinary business people to imitate?

Not at all. Even if you can't match the completion speed, you can start imitating the way the process is divided today.

The Secret of Genius is Passing Your Own "Template" to AI

The most helpful aspect of Mr. Tanaka's method is that he does not let the AI think from scratch.

Mr. Tanaka explains that he has the AI learn the analysis and material templates he has used in his work, then inputs research results and his own hypotheses to have them converted into the necessary format.

In other words, the division of labor between humans and AI is as follows:

What humans handle: Purpose, hypotheses, judgment, experience, intuition, final confirmation

What AI handles: Transcription, classification, summarization, reordering, translation, format conversion

This boundary line achieves both speed and quality.

A common failure is to dump everything on the AI with a request like "make a nice proposal document."

This results in something that looks neat but has thin content. The AI cannot see what the points of contention are in that specific company. It doesn't know where the boss might object or what the customer is afraid of.

Mr. Tanaka's method is the opposite.

The human outputs all the specialized knowledge and hypotheses in their head via voice. The AI then converts that material into logic it can read. Polishing the appearance comes even later.

The secret to speed is not just input speed. It lies in separating "thinking," "arranging," and "showing."

STEP 1: Output Materials via Voice Before Opening PowerPoint

The first thing to do is not slide creation. Speak the materials that should go into the document.

You don't need to speak in beautiful sentences. It's okay to correct yourself halfway or for the order to be jumbled. The important thing is to provide judgment materials that the AI cannot supplement on its own.

At a minimum, please speak about these six points:

• Who are you showing this document to?
• What do you want the recipient to decide?
• What is the main conclusion you want to convey?
• What are the facts and figures supporting the conclusion?
• What are the expected counterarguments and risks?
• What do you want the recipient to do next after seeing the document?

For example, for an internal new project, speaking like this is enough:

"A document to get the sales team manager to approve the trial introduction of a new customer management method. The conclusion is to try it for 4 weeks in some teams."

"Currently, recording methods differ by person, leading to missed handovers. Counterarguments are increased input work and the possibility that the field won't use it."

"I want to narrow down the input items and decide on continuation after the trial."

At this stage, don't think about design. Don't even arrange the order. Output as many materials as possible that are necessary for judgment.

STEP 2: Ask Claude for "Issue Organization" Instead of "Summarization"

Once you have transcribed the voice, don't turn it into slides immediately.

First, break it down into purpose, claims, evidence, counterarguments, and missing information. Finding logical holes here prevents the document from becoming something that only looks impressive on the outside.

The following is not a direct quote of the prompt published by Mr. Tanaka. It is a reconstruction of his published workflow into a form that can be used for general material creation.

▼ Copy from here

You are an editor who designs decision-making materials.

Below are the materials for a document I spoke via voice. Do not make them into slides yet; instead, organize the content in the following order:

1. Reader of the document
2. What you want the reader to decide
3. Key message serving as the conclusion
4. Evidence supporting the conclusion
5. Expected counterarguments and risks
6. Missing facts, figures, and items to be confirmed
7. Next action requested of the reader

Rules:

• Do not create facts or figures not in the input.
• Do not guess ambiguous parts; write "To be confirmed."
• Point out where logic is missing and explain why.
• Group similar content.
• Separate content with weak relevance to the conclusion into "Candidates for deletion."

Voice Transcription:

[Paste transcription here]

▲ Up to here

Once you receive the output, the human fills in the "To be confirmed" and "Logical leaps."

This extra step is effective. This is because AI can organize even incorrect content to be easy to read. Confirm figures, citations, proper nouns, and causal relationships before making them into beautiful slides.

STEP 3: Convert to a One-Message-Per-Slide Story

Once the points are gathered, next, create the order of the slides.

The trick here is not to make the headings "item names."

Headings like "Background," "Issues," and "Proposal" alone don't tell you what the slide is trying to say. Make it so the conclusion is conveyed by the heading alone, such as "Customer information is dispersed, increasing confirmation work during handovers."

▼ Copy from here

Please convert the previously organized content into a slide structure for decision-makers.

Purpose: To make a document that allows for understanding and judgment in a short time.

Number of slides: Aim for 8-12 slides, adjusting according to the content.

Please output each slide in the following format:

Slide Number:

Message Heading: The conclusion to be conveyed on that page in one sentence.

Body: Up to 3 key points supporting the heading.

Evidence: Necessary figures, facts, and citations.

Diagram Plan: Comparison, time series, causal relationship, procedure, etc.

Confirmation Items: Parts where evidence is insufficient.

Structure Rules:

• One message per slide.
• Show the conclusion, reason, and requested judgment in the first 3 slides.
• Do not hide counterarguments and risks.
• Do not add figures or facts not in the input.
• Do not repeat the same claim in different words.
• End with the next action.

▲ Up to here

The purpose of this prompt is not to shorten the text.

It is to rearrange it into an order of "what the other party needs to know to make a judgment."

If it's for an investment committee, the points would be project background, profitability, downside risks, value enhancement measures, and exit conditions.

If it's a sales proposal, it's the customer's current situation, the inconvenience if left alone, the proposal content, the basis for introduction, and the execution procedure.

If it's an internal project, it can be replaced with purpose, current situation, options, recommended plan, risks, and items to be approved.

Even if the industry is different, it can be applied by centering it on "who and what to have them judge."

STEP 4: Convert to Marp Format with Claude

Only after the story is solidified do you change it to a slide format.

Marp is an open-source framework that can generate slides from Markdown-formatted text. According to the official website, it can output to HTML, PDF, and PowerPoint.

Source: Marp Official Website

https://marp.app/

The reason Claude and Marp are combined in Mr. Tanaka's demonstration is clear.

It's because slides can be treated as editable text rather than a collection of shapes. You can collectively instruct changes in order, heading corrections, and page additions.

▼ Copy from here

Please convert the following slide structure into Markdown that can be displayed in Marp.

Settings:

• theme is default
• Display page numbers
• 16:9 aspect ratio
• One message per slide
• Headings should be short and convey the conclusion
• Body should have a maximum of 3 items per slide
• If there is a lot of text, divide the pages instead of making the text smaller
• Leave "Diagram Plan:" as a comment on pages where a diagram is needed
• Do not add figures, facts, or citations not in the input
• Display "To be confirmed" for parts where evidence is unconfirmed

Slide Structure:

[Paste output from STEP 3 here]

▲ Up to here

If you open the outputted Markdown in something like Marp for VS Code, you can check it as a slide.

However, it is not complete at the point it is generated. Adjust the company-specified templates, fonts, color schemes, logos, and handling of confidential information according to your organization's rules.

The Faster You Create, the Heavier the Human Confirmation Becomes

The most dangerous thing in AI slide creation is that mistakes are neatly organized.

If the sentences are natural and the layout is tidy, the content also looks correct. However, AI may supplement figures not in the input or create strong conclusions from weak evidence.

Before completion, check at least the following items:

• Do the figures match the original materials?
• Do the citations exist, and has the context not changed?
• Are facts and hypotheses not mixed?
• Are the axes, units, and periods of graphs correct?
• Have you not entered customer names, personal information, or undisclosed information into external services?
• Are the risks and counterarguments of the proposal not hidden?

Confidential information requires particular caution. You must not enter customer information or undisclosed financial information into AI services that your organization has not permitted. Use anonymized materials or an environment approved by the company.

Mr. Tanaka's method is not a technique for dumping judgment onto AI.

Humans provide the materials for judgment, AI handles the conversion, and finally, humans verify the facts. That's why it works as a job, even if it's fast.

If You're Applying This to Ordinary Slide Creation, These 4 Steps Are Enough

Even if you don't have the same investment experience as Mr. Tanaka, you can imitate the process.

1. Speak the purpose and claims of the document for 3-5 minutes via voice.
2. Pass the transcription to Claude and organize the points and missing information.
3. Create a one-message-per-slide structure and convert it to Marp format.
4. Humans confirm figures, citations, confidential information, and final judgment.

You don't need to create a perfect system from the start.

Choose just one document you're making tomorrow and speak for 3 minutes before opening PowerPoint.

Who are you showing it to?

What do you want them to decide?

What is your conclusion?

What is the evidence?

Where might they object?

Paste that transcription into the first prompt.

The essence of Kei Tanaka's genius AI slide creation method is not just using Claude or Marp.

It is about putting the judgments that only humans can make first and connecting the conversion work that AI is good at behind it.

If you want to speed up slide creation, you don't need to rush your PowerPoint operations.

Change the order in which you create. Start from there.