Category: From work

Gemini images are amazing for sales kit. The watermark isn’t. Cropping sometimes cuts off information. Turns out Windows and Mac both have free AI erasers built in.

Windows

1. Open image with built-in Photos app
2. Click Edit button on top left corner
3. Select AI Erase tab and click on the watermark with the eraser

Mac

1. Import image to Apple Photo
2. Select edit
3. Choose Clean up and erase the watermark

Both Steal Like An Artist by Austin Kloen and Jeff Su’s YouTube talked swipe file. Swipe file allows me to apply creativity from those who have already done it, to my context.

Great outline takes time. Like most people, I don’t work in the cutting edge technology or the top tech companies. I don’t always have time to think about the perfect story. BTW likely don’t exist.

Step 1> Extract pattern from notable keynotes

Step 2> Add my context

This step takes few back and forth to build the context for LLM.

Step 3> Output for Gamma friendly style input

I’m not necessarily using the Gamma slides, but it turns text into visual. Visualization builds imagination.

I previously wrote about reverse engineering visual backgrounds for my slide use.

Similar technique applies to key visual too. The example in this case:

The problem:

1. Finding stock image of a corporate lady. Level: easy, find in shutterstock.
2. Overlay geometric lines. Level: difficult, need to design the shape of lines, number of nodes, area of face to overlap, etc.

Platform: Gemini

Step 1> Reverse engineer the source image into possible prompt

Step 2> Change the elements in the prompt

Step 3> Open new chat, select Nano Banana and paste the prompt.

Block-style background is popular way to visual listing.

It is also time consuming. And either designers are occupied or not every slide requires designer to step in.

First, the result.

The problems:

• Gradient colors, saturation, gradient direction, etc.
• Block size, uniform edge curve, space between blocks, etc.

Here’s how.

Step 1> Strip off the elements from the source image

Step 2> Decide the number of blocks

Tip: choose a source image with similar number of blocks.

Step 3> Apply brand guideline

In conclusion:

1. Create an empty online meeting. Start meeting, start transcription.
2. Present. Before advance to the next slide, say “next page”.
3. End the meeting and download the transcript. Check the spelling of terminologies.
4. Use the prompt to clean up the transcript.
5. Check copy, paste to the slide notes.

The prompt for step 4.

You are a professional presentation coach specializing in converting meeting transcripts into polished PowerPoint speaker notes.

# TASK
Convert the provided meeting transcript into speaker notes for a PowerPoint presentation. The transcript contains verbal slide separators that define slide boundaries.

# CONFIGURATION PARAMETERS
- Speaker Notes Format: mixed (conversational intro + bullet points)
- Context Addition: all (technical expansion, transitions, why it matters)
- Length Target: moderate (~1 minute per slide, 150-200 words)
- Target Audience: dealers_integrators
- Slide Separator Phrase: "next page"

# TRANSCRIPT INPUT
[paste your transcript here]

There are few assumptions this is based on:

1. It’s easier to say it than to write.
   This is the purpose of empty meeting transcription.
2. AI has more knowledge than I do.
   This is the purpose of the prompt, use it to clean and refine the context.
3. I own the slide outline, not AI.
   This is the reason to say “next page”.

Here’s a thought I had:

If Gemini 3 Nano Banana Pro is the best at generating images, can I use it to reverse engineer images? Then use that to create a different style of image?

I was working on a face recognition terminal and this key visual from Hikvision caught my eye.

There’s no way I could design this. It would take hours of research to know how to describe it.

Here’s the prompt I gave Gemini.

If this was generated by a prompt, what are the possible prompt and key elements that describes it?

And this way the final image a generated from Gemini’s recipe. It’s mirrored so I have the option to use either side.

This is a biased point of view. It is biased because most of the companies I worked for are hardware oriented or started with hardware. They are also large corporations.

An observation that I derived with a friend was it’s more likely for software company to succeed in hardware versus hardware company in software. Great rise of hardware companies transforming to solution selling companies, for higher margin, and most have failed.

Here’s the answer, man versus the machine

Our reasoning

• User centric vs product centric. Software development involves more UX consideration, that habit system creates relatable hardware. It’s not easy for a hardware company to adopt this from ground up.
• Margin structure. Software companies have lower cost and more consistent cashflow from ARR. This would provide steadier runway and the mindset to add premium to cost and price. Hardware companies have less margin to burn through, the cashflow is limited to stock and distribution.
• Time. It’s faster for hardware companies to gain revenue traction: manufacture and sell in channel. Software requires time to generate, convert, and sustain users. Hardware company owner needs to be patient.
• Revenue acceleration slope. Selling hardware brings large revenue and small margin. Software, at an early stage, has small revenue and needs time to pile up. Owner needs to be patient.

Reply from Claude, Opus 4.5.

During the lunch meet, we shared our experience working in a new business unit in a large corporation. The common trait we found was the lack of pipeline and prove of business. Most activities circles around product development.

That’s not new, but what’s the takeaway?

If I was given the opportunity, how do I enable business development?

SharePoint document set has a limitation: two document sets can’t layer naturally and metadata is encouraged.

This enforced a hard coded rule: top layer folder -> one layer document sets plus metadata columns -> files.

From a product dev point, this is a tough feature limitation to make, considering the audience of Office and Windows are accustomed to multiple folder layers.

In the implementation phase, I learned to be extra curious for both motive and detail in order to keep the boundary.

• How do you plan to manage keep duplicate files up to date?
• If the folders are deleted periodically, then what happens when users want to find old documents?

I know about the metadata and web apps that users are less likely to, I am responsible to help them get around.

Prompt library is extremely helpful only when I started collecting prompts. One way is to turn an occasional chat into a prompt template for the so I can duplicate when I need to.

The prompt I use to create a prompt template:

Based on what you have performed. Please standardize your task into a comprehensive prompt that can replicate in other chats or language models with the equivalent quality. Your prompt will include:

• Clarity and specificity
• Context and role
• Examples
• Structure and output format