Hello and welcome (back) to The Mindshift AI Inference!

Most newsletters disappear into the inbox. They often never get read. Their informational value is lost forever.

Unless you give your AI agent a way to uncover it for you.

I am preparing for my annual MiCANNES AI panel and tried a simple experiment yesterday:

I took one newsletter source, Superhuman, and treated it not as email, but as a dataset.

The Workflow

I did not do this manually.

I instructed Codex, the AI agent working inside my second brain, to search my inbox for all recent emails from the newsletter sender.

First, Codex used the gog CLI to search Gmail for all recent Superhuman newsletters.

Then it saved each newsletter as a plain markdown file. Nothing fancy. No proprietary database. No dashboard. Just text files in a folder.

If something is in markdown, I can easily hand it to an AI system and still understand what is happening.

The next step was structure.

Newsletters are not random. They have recurring sections: Today in AI, productivity tools, research updates, weekend features, robotics, science, media, and so on.

Once you see that structure, the inbox becomes less like a pile of messages and more like a source library.

What becomes possible now: “Give me all ‘weekend features’ of the last two months”, “Has newsletter x written anything about topic y recently?”

The Old Way And The New Way

OLD WORLD: You remember that you saw something about AI music in a newsletter. You search your inbox. You click through emails. You skim manually. You lose patience.

NEW WORLD: The newsletters are saved as text. A small script extracts the relevant sections. Search finds every mention of music, audio, piano, songs, lofi, or orchestras. AI helps summarize the findings.

The difference is not the AI.

The difference is the architecture.

AI is useful when the underlying material is accessible. It becomes much less useful when the material is trapped inside apps, inboxes, formats, and half-remembered tabs.

What We Found

Once the newsletters were saved, I started asking Codex questions.

First, I asked it to extract all items in the “Today in AI” category from one month of newsletters.

Then I asked it to use the same saved files to build topic reports. One report covered AI consulting. Another covered PowerPoint generation. Another covered talking AI humans. Another covered Meta opening its ad ecosystem to third-party AI tools.

Finally, I asked it to search the whole archive for music-related items.

That surfaced AI song generators like Suno, tools that generate music alongside other media, Anthropic’s lofi study music channel, robotic hands playing piano, Sophia performing with a symphony orchestra, and even a modular audio system for vinyl, CDs, and cassettes.

None of this required a complex knowledge platform.

It required a folder, markdown files, search, and a small parser.

The Bigger Point

This is the pattern I keep coming back to:

If your information lives only inside someone else’s interface, you do not fully control it. You can use it, but you cannot easily build on it. I wrote about this in Building Your Own Creativity System: the system only becomes powerful when the context is accessible and under your control.

If your information lives in plain files, everything changes.

You can search across it.
You can extract patterns.
You can create reports.
You can write scripts.
You can ask AI to reason over it.
You can keep the system yours.

The future of knowledge work will not be one perfect AI tool. It will be small, understandable systems that connect the information we already have with the intelligence we now have access to.

The inbox is not the destination anymore.

It is raw material.

Have a great day!

Matthias