Hi AInauts,

Welcome to the latest issue of your favorite newsletter.

You know those small subscriptions that quietly charge every month because they solve one tiny job well enough? We know that feeling too. Ouch.

That is exactly where AI is getting interesting. More and more often, you can have small tools built for your own workflow instead of renting yet another SaaS. Today we are looking at those mini-apps, ChatGPT as a real workspace, and Germany's AI location index.

Here is what we have for you today:

Let's go.

The president of OpenAI. The CEO of Google. The founder of LinkedIn. We talk to the people building AI. Every morning, we put what they told us in your inbox.

What they're actually worried about. What they're betting on. What's coming next. Delivered in five minutes, before your first meeting.

With one email, and only five minutes, you can stay ahead of 99% of the world.

Join 75K Readers →

Ralf killed a SaaS inside our AI Automation Community. A completely normal work moment for him, which is exactly why we like the example so much.

For meeting notes, he could have paid for another subscription tool. Instead, he built and shared HyperScribe: a small desktop app for macOS and Windows that uses Deepgram to transcribe audio, separate speakers, and do the job for roughly 30 cents per hour.

The next day, he followed up with his "Confessions of a SaaS hitman." With HyperRec, he built a small tool that records microphone and speaker audio directly. Another subscription replaced by one focused workflow.

The motto: tool annoys you, subscription annoys you, agent builds a better shortcut. Done.

We say this so often now that we are starting to annoy ourselves. Sorry. But the point matters.

You do not need to know how to code. The technical part is secondary. You need a job that you can describe.

The real blocker is rarely the tech. It is the moment where you dare to give your agent a real job.

And if the agent claims it cannot do it, get a bit stubborn: "I know you can do this. I believe in you."

Not a joke. Sometimes that kind of push is exactly what helps.

Second example before we build one ourselves: last week I visited an old friend from my teenage years. Martin works as a building services planner and engineer. IT, programming, and vibe coding are not exactly his natural habitat.

Our starter prompt: "I am here with Martin. He works as a building services planner and engineer. I want to show him what is possible with Codex. Give me 10 ideas for a small mini-app that would be relevant to his work."

A few minutes later we had the idea: a pump pre-selection tool that suggests suitable products from relevant manufacturers based on project constraints.

A bit of prompt ping-pong later, the first version worked. Then we asked:

More than half of the suggestions were implemented while we sat in the garden talking about old times and future plans.

With a few clicks, we deployed the app for free on Cloudflare Pages, password-protected and ready.

On the way home, his WhatsApp arrived. In plain English: I can actually use this. I got a taste for it. That is exactly what you want from a first prototype.

If you want to start, start small. Our first "app" was a simple browser extension, painfully prompted with ChatGPT back then. Since then we have built dozens of small and large tools with AI.

You still do not need to have built an app before. Install the Codex App from OpenAI or the Claude Code / Claude Cowork desktop app as your base.

Tip: you can also ask ChatGPT or Claude first, because that chatbot probably knows you better than a freshly installed Codex or Claude app. Then add: "IMPORTANT: write this so that I can paste it directly into Codex / Claude Code."

The most important sentence is not at the top. It is further down:

Do not write code yet.

Unsexy, but decisive. Many agents sprint off, build half an app, solve three side problems, and forget why you started.

This prompt forces the agent to understand your situation first and save the context in files. PLAN.md and TASKS.md are not bureaucracy. They keep the agent oriented after 40 minutes.

Anthropic recommends exactly this kind of clean task split in its agent guides: clarify goal and acceptance first, then automate. In practice, that is the difference between "looks cool" and "still works tomorrow."

While testing the prompt, we built a small app that turns a webinar transcript into working materials for us: summary, infographic prompt, mailing idea, and course notes.

If you have a button after two hours that takes a real file and produces a useful output, you have learned more than from ten tool demos.

Maybe you cancel the subscription afterward. Maybe you keep paying for it, but finally for a good reason.

Real apps need more later: security, data, tests, hosting, support, payment, and the boring parts that keep things alive.

For today, aim smaller: build one tool that solves one real problem in your day. Then take the win.

You built something while others were still collecting examples.

P.S. If you need more inspiration, this X post includes twelve prompts where the author replaced smaller AI tools, from grammar checks to meeting summaries and SEO briefings.

Most AI news is a waste of time. The Future Today is a daily 5-minute read focused on what actually matters.

You’ll get exclusive interviews with the CEOs, researchers, and builders shaping AI. Plus top stories broken down simply and practical advice you can immediately apply.

Read the newsletter trusted by teams at NVIDIA, Google, Anthropic, Meta, Dell, and Salesforce.

Join Free Today →

OpenAI is pushing ChatGPT and Codex in the same direction: away from the pure answer window and closer to your actual work.

The effect may matter more than the next model rumor: you have to trigger fewer routines manually.

On the new Scheduled page in ChatGPT, you can see scheduled tasks, check the next run, pause, resume, edit, or delete tasks.

Tasks can handle reminders, recurring work, and monitoring. Less "ask me again later." More "I will tell you when something happens."

Daily briefings, email monitoring, concert alerts, sale alerts: ChatGPT can watch the things you used to check manually.

If you are not using Scheduled Tasks yet, click here and create one. Takes five seconds.

OpenAI even removed the popular Premium feature Pulse. We read that as a clear signal: the chat should stop only answering and start taking on small obligations.

The smaller ChatGPT updates point in the same direction and make the chat window feel more like a work environment.

The new Tasks page is currently meant for ChatGPT on web and mobile, not the desktop app or Codex. That detail matters, because OpenAI is also moving Codex closer to the computer. See our earlier post. The relevant pieces are:

Many productive AI workflows are not blocked because the model is too dumb. They are blocked because the work lives in other tools and surfaces: Gmail, LinkedIn, Salesforce, WordPress, internal admin tools, local apps.

More Codex features have started rolling out for us over the past few days. That is where things get interesting. And risky.

If Codex can use Chrome, it can see logged-in pages. If Codex uses Computer Use, it can click, type, and operate windows. Powerful, yes. Also not a casual prompt.

Record & Replay is next! The idea is brilliant: you show Codex a recurring workflow once, and Codex turns it into a reusable skill.

Still, the direction is clear: from "describe the workflow" to "show me the workflow once." That is much closer to how normal people work.

Outside pressure explains why OpenAI wants to consolidate the interface. ChatGPT is still huge, but Sensor Tower data says its former 90 percent share has fallen to 46.4 percent of the global AI-assistant market.

Gemini sits at 27.7 percent, Claude at 10.3 percent. ChatGPT is still enormous with more than 1.1 billion monthly users. The market is not collapsing; it is growing and spreading out.

Gemini is attached to Android and Workspace. Claude keeps growing with power users. Microsoft is rolling out Copilot Cowork globally for Microsoft 365 Copilot customers.

Agents are becoming small workflows with a budget line.

For your daily work, the decisive question is not whether ChatGPT, Claude, Gemini, or Copilot is ahead in a benchmark. The question is: which agent can you give a job to without feeling nervous afterward?

Waiting is not our favorite strategy. Start with these features:

Have fun with it. Ideally, it gives you a real productivity bump.

Berlin is Germany's strongest AI location, at least according to the new KI-Standort-Index Deutschland 2026.

Berlin lands in first place with 84.7 points, ahead of Munich at 77.9 and Stuttgart at 66.1. After that, the gap gets steep: Hamburg reaches fourth place with only 43.3 points.

The ranking compares 37 cities across four pillars:

So far, so measurable.

But what is an AI location, really?

A few lighthouse projects are not enough. A city becomes relevant when several things happen at once: universities train AI people, companies hire them, startups build products, and people can afford to live there.

We think the ranking is a useful reality check. Berlin, Munich, and Stuttgart are not at the top because they make the loudest AI promises. They are visible because more of the system already exists there: jobs, research, startups, and economic pull.

The German question becomes European very quickly. Even the best location remains dependent when models, chips, cloud access, and platform rules are controlled elsewhere.

The bigger picture is this DW video: US limits AI access - What it means for Europe. It asks whether AI access itself is becoming a geopolitical tool.

P.S. If you want to hear how one of the most important American AI CEOs currently sees the world, the long interview with Anthropic CEO Dario Amodei is a good bonus: here you go.

Done. Today's issue had a lot of agent future, and a lot of agent present, but the practical assignment stays grounded: pick one recurring job you can describe cleanly. Let an agent plan first. Then let it build.

And if one subscription disappears along the way: even better.

See you in the next issue.

Reto & Fabian from AInauten