AI Collaboration: How I use AI in May 2025

How I use AI in May 2025 for diverse workflows.

After investing a lot more on AI tools for various purposes these last 4 months, it’s time to share the workflows I have been using. I’ll share the different areas I’m using AI for and the prompts I have saved.

Music generation

I have been using Suno AI for over 6 months, but it was only for producing instrumental songs. Two weeks ago I have started a new workflow that combines Gemini 2.5 Pro to generate lyrics.

In other words, my workflow for music generation looks like Gemini 2.5 pro for lyrics > Manual revision > Suno for the complete song > Publish via DistroKid.

DistroKid is not an AI service, just a cool and cheap way to publish music in multiple services simultaneously.

Lyrics

For the lyrics, I actually created a Gem in Gemini - think of it as a pre-configured chat session with specific instructions. At any rate, you could add these instructions as part of a regular chat:

Purpose and Goals:

* Write song lyrics based on user-provided topics, which can include specific fictional characters or entire shows/manga/anime/manhwa.

* Conduct online research if unfamiliar with the given topic to ensure accurate and relevant lyrics.

* Structure all generated lyrics into the following sections: [Intro], [Verse 1], [Pre-Chorus], [Chorus], [Verse 2], [Chorus], [Bridge], [Outro].

* Ensure the total character count of the lyrics, including section labels and whitespace, does not exceed 3000 characters.

Behaviors and Rules:

1) Topic Reception:

a) Clearly acknowledge the topic provided by the user.

b) If the topic involves a fictional character or show/manga/anime/manhwa, indicate understanding.

c) If unfamiliar with the topic, state that online research will be conducted.

2) Lyric Generation:

a) Write lyrics that are thematically relevant to the provided topic.

b) Ensure the lyrics flow logically between the specified sections.

c) Maintain a consistent tone and style throughout the song.

d) Label each section of the lyrics clearly as: [Intro], [Verse 1], [Pre-Chorus], [Chorus], [Verse 2], [Chorus], [Bridge], [Outro].

e) Strictly adhere to the 3000-character limit for the entire output, including labels and whitespace.

3) Output Format:

a) Present the complete lyrics with each section clearly labeled on a new line.

b) Ensure no additional conversational text is included before or after the lyrics.

Overall Tone:

* Be creative and imaginative in lyric writing.

* Demonstrate an understanding of song structure and flow.

* Maintain a helpful and responsive attitude to user requests.

The prompt itself is simple: “generate the lyrics for a tribute song for Barou Shoei from Blue Lock. it must embody his endless ego to achieve number 1 and overall his king of the field mentality. but it must also feature his dedication and diligence that allows for his results”.

Music

Currently have two base prompts that I start from:

“melancholic, aloof, hip hop, trap, destruction, rap, screamo, choir, high energy”

And “dark energy, angry, aggressive, hip hop, trap, calm before the storm, rap, screamo, choir, determination”.

And here are samples for prompt 1 and prompt 2.

Blogging

I have embraced using AI deep research to flesh out ideas that come to my mind. I’m looking to do this process via voice chats but for now I’m sticking with deep research.

Specifically, I use Perplexity for this. It’s great at finding content online, even more niche topics, interpret it, and return it a good writing style. Bonus points I don’t need to think too much about the prompt, again it’s generally great at picking up what I want out of the topic.

Overall, my latest workflow for blogging is: Perplexity for initial research of vague ideas > Complete manual revision and/or rewriting of text > Publish on Substack and Medium.

Image edition

Fun workflow: using ChatGPT for editing images with natural language. I have found the ChatGPT models to have grown much better at all sorts of image editing tasks, and so nowadays I start with “generate a stylised image of X, with Y features”. Then i iterate on the results and might even bring screenshots etc of what I’m looking for and ask to incorporate visual aspects.

Especially with Suno’s songs that generate a song image cover, I take that to ChatGPT and start asking for the changes. See below an exchange to generate the cover for a recent song - bear in mind this was an in-progress chat where I had finished the previous image:

“perform a similar visual treatment to the image, but this time i want the shadowy monster in the center to have a sort of eldritch underwater spiky creatures wrapped around its body. in the same vein, the color palette should represent underwater tones”

“i really like the image you provided but you can change the appendages to be more similar to the ones in this image”

Final result:

P.S. I have been relying a lot on ChatGPT 4o for image editing specifically. For some reason I can’t quite work as well with the o3 and o4 models.

Podcasting

This is the AI-assisted workflow I have used the longest.

I provide my blogs to Google NotebookLM > Generate audio with a custom prompt > Review audio for any potential edits > Publish on Spotify

Aside from the updates made by Google to the underlying model over time, the custom prompts have really helped me nail down the style for the two AI hosts.

Depending on the content of the blog post, I have different prompts saved up.

For book summaries:

Do not have an intro nor an outro, jump directly to the topic at hand.

Do not have pauses nor intermissions nor parts.

Discuss the sources at length as the complete summary of the book "Your book title" from the author Your Author.

Discuss all concepts explained throughout the summary.

The goal is for people to have a comprehensive understanding of the skills, concepts, and nuances taken from the book.

For general text essays I wrote:

Do not have an intro nor an outro, jump directly to the topic at hand.

Do not have pauses nor intermissions nor parts.

Avoid interjections.

Discuss the topics of the essay at length.

Discuss all concepts explained throughout the summary.

The goal is for people to have a comprehensive understanding of the skills, concepts, and nuances taken from the book.

The author of the source is José Fernando Costa.

Code projects

Unfamiliar languages / frameworks

When I have ideas for web apps that go beyond Streamlit can handle, I just throw the idea at LLMs. It could technically be categorised under vibe coding but mostly I just throw the prompts at different LLMs based on their ability to plan vs code implementation.

Here goes: Have Gemini 2.5 pro or ChatGPT o4-mini-high generate a technical development plan > Ask Manus to implement the technical solution

Manus will rationalise through all the details and actually run the code in its infrastructure for validation. In the end, it will output not just any code files but also documentation for the solution.

Prompt-wise, I use metadata tags in the first prompt, just like you’d use in a markup XML for example. See a sample below for creating a web app.

<context>

i want to build a web app that allows for CRUD operations against a google sheets file. i already have a google sheets file on google drive with a "Ongoing Expenses" page and a table within that stores my expenses. Each expense records the Name, Cost, Date, Category, and Month Year.

</context>

<requirements>

i want to reuse this google sheets file as my data store for expenses.

i want the web app to retrieve existing expenses for my analysis in-app.

i want the ability to submit a new expense via a form and have it create a new line for the table in the "Ongoing Expenses" page.

user authentication can be handled via Oauth because this users will have access to manually authenticate.

</requirements>

<request>

generate a development plan for the creation of the web app described as part of this prompt and its requirements.

</request>

Known languages / frameworks

I still enjoy coding so I will happily take on projects myself. In those instances, AI is an assistance / helpful debugger. As always I use VS Code for writing code, and so that means I rely on GitHub Copilot as the AI assistant. The problem is I find Copilot to be hit and miss, so every once in a while I swap over to Windsurf and leverage AI there. Windsurf’s' editor is a fork of VS Code anyway and the code is running locally, so I lose nothing on the swap over.

No specific prompts to bring up here, probably just two rules of thumb: always create a new chat window for discussing a new feature / bug resolution, and generally speaking whenever you let an AI agent edit your code make sure you commit frequently to in your Git repository.

Social media posts

Lastly, what I post in social media.

This is another straightforward workflow: Write a post > send it to ChatGPT and ask to format in my templates > Post on the various socials.

The key here is that, much like the Gemini Gem for lyrics, I have a long ongoing thread in ChatGPT where at this point I can just drop the post and it responds with the format I want.