Hi AInauts,

Welcome to a new edition of your favorite newsletter!

Last week, we mentioned in passing that we had replaced Wispr Flow with our own app. We received more messages about that one sentence than almost anything in recent memory.

Nearly everyone asked the same thing: “How? And can I have it?”

Yes, you can. But today's issue is really about something bigger.

Anyone can build now. Or generate: images, voices, videos, and even someone to talk to on a live video call. The tools are no longer the limit. The real limit is whether we come up with smart ways to use them.

Here is what we are covering today:

Let's get started!

AI is moving fast and most people are falling behind. 

The Rundown AI keeps you ahead of the curve. 

It's a free AI newsletter that keeps you up-to-date on the latest AI news, and teaches you how to apply it in just 5 minutes a day.

Plus, complete the quiz after signing up and they’ll recommend the best AI tools, guides, and courses — tailored to your needs.

Sign up to start learning.

A quick reminder: Wispr Flow is a dictation app. Hold a key, speak, and release it. Cleanly written text appears wherever your cursor is. We loved the app, used it every day, and dictated more than a million words with it.

Today we use our own version: AInauten Voice.

It does exactly what we need:

You have two options: Download the installer from voice.ainauten.com and install it directly on your Mac. Or use the source code and let your agent install the app for you.

We recommend the second route. We have only tested the app on our own Macs, so your agent can inspect the code first and adjust it if anything causes trouble.

Use the ChatGPT or Claude desktop app for this. Do not use a normal chat. Use Work or Codex mode in ChatGPT, or Claude Code.

Those environments can create files and build programs on your computer. You will generally need a paid subscription.

Option 1: You have a Mac with Apple silicon. You need an M1 or newer, macOS 14 or later, at least 8 GB of RAM, and about 3 GB of space for the speech models.

Give your agent this prompt:

A setup assistant then walks you through the rest: downloading the models, importing Wispr Flow settings, granting permissions, and running a test dictation. Wispr Flow is only closed, not deleted, so you can switch back at any time.

You should never install an app that listens and types into other programs without inspecting it first. That includes ours. This is why the prompt says: inspect first, then install.

Option 2: You use Windows, an older Mac, or want your own version. Ask your agent to build one:

Is the app perfect? Probably not. It is a weekend workshop project, not a polished commercial product. It works for us, but you may still hit a snag. If that happens, send your AI a screenshot and ask it to fix the problem, or send us feedback through the app.

The app did not appear in 30 minutes, especially because the agent needed time to write the code. But the process was not difficult.

The third story below shows you how to build your own app.

Here are a few practical tips that are useful well beyond coding:

Tip 1: Put feedback in the queue. You do not need to wait until the AI has finished before giving feedback. But you also do not need to interrupt it and push it in a new direction.

We like ChatGPT's queue feature. In the desktop app, open Settings, then General, Composer, and Follow-up message behavior. Choose Queue.

The AI then processes your feedback in order. To send one message immediately, hold the Command key while sending or use the arrow's Steer option.

Some people even keep their system running overnight by queuing several “Continue and improve it” instructions. That can work, but it may also produce surprises the next morning. 😁

Tip 2: Use the right model for each step. We use the strongest model for the specification and review, such as GPT Astra or Claude Fable/Opus. A faster model, such as GPT Sol or Claude Sonnet, is usually enough for the actual implementation.

This is not just our intuition. Anthropic recommends the same pattern in its Advisor strategy, and Claude Code even offers an opusplan setting for it.

Tip 3: Run the work as a goal. An AI normally stops when it decides it is finished. Often, the app is only half built at that point. A goal lets you define what finished actually means. Use the /goal command.

Before stopping, the AI checks whether the goal has been met. It keeps cycling through build, test, and fix until every requirement passes or it reaches a decision only you can make. This produces better results, but it also uses more tokens.

P.S. We understand that dictating out loud is not ideal in every environment. But there is a solution for that too, at least as long as nobody is looking directly at you. 😁

P.P.S. Yes, that is in our app as a beta too. Unfortunately, it does not work in German yet.

Parallel, Turbopuffer, and Wordsmith run their entire GTM motion on Attio, with agents that chase every buying signal, build pipeline, and move deals forward, 24/7.

Try Attio for free

We have seen a lot by now. But this genuinely impressed us.

A few days ago, Tavus introduced Griffin, an AI counterpart for live video calls. This is not a prerecorded video. Griffin listens while speaking, nods, says “mm-hmm,” and does not interrupt just because you pause to think.

All it needs is one photo and about ten seconds of your voice.

Watch it here.

Fifty-four people were sent into a one-minute video call. They were told they were speaking with another test participant. Only at the end were they asked whether the person on the other side might have been AI.

Twenty-six of 54 were convinced it was a human. That is 48%.

For comparison, Tavus's previous system fooled only one out of 41 people in the same test. This was Tavus's own study, the sample was small, and the call lasted only one minute. But watch the videos and you will immediately see how realistic they are.

Live avatars already existed, including HeyGen's version. Griffin still feels like another level. Tavus is therefore releasing the first version only to selected testers. We applied and will report back if we get access.

Griffin did not arrive alone. Several releases in the past few days have made AI output even more realistic:

These tools now cost almost nothing, and your agent can set them up for you. You still need to decide what they should be used for. That brings us to the next story.

To finish, here is an idea that has been on our minds over the past few days.

We used to learn software. Which button does what? Where is the feature in the menu? Traditional software has a feature list. You can count the items.

AI does not have a feature list. It is an open field. You have to imagine what it could do for you.

Ralf from our community put it well last week. He had given an introductory AI session, and nine of the 11 participants had barely used AI before.

One question dominated at the end: “So how exactly do I use AI in my own workplace?” His conclusion: “For the first time, mainstream software expects people to imagine the features themselves.”

There is something to that. We were trained to be users. If the tool cannot do something, it cannot be done. A builder asks a different question: not “What can this tool do?”, but “What could I do with it?”

That is exactly how our dictation app came to exist. We did not search for the right tool. We built one. You can do the same without programming skills, often with the ChatGPT subscription you already have.

You may be thinking: “I already build my own apps.” If so, well done. You are already a builder and far ahead of many people around you. Take the prompt below anyway. It might contain your next idea.

For everyone else: You are one prompt away from joining them. Give the first step 15 minutes. The AI can build the rest.

Rule of thumb: Start with the smallest format that solves your problem.

The sequence matters. Start with the idea, then create images or mockups, then write a clear description, and only then start building.

Paste the following prompt into ChatGPT Work/Codex, not a normal chat, or Claude Code/Cowork in the desktop app, also not a normal chat.

Use the best available model to begin. You can switch to a faster model for the implementation later.

Do not give up too soon. The first attempt is almost never very good. Give feedback, take a screenshot, describe the current and desired result, and start another round.

One warning: It has never been easier to build something nobody needs. Ask yourself: useful or merely new?

Do not let that stop you. If it saves you time, money, or frustration, you have already won. And once you build something for the first time, an entirely new universe of possibilities opens up.

Tell us what you built. We are curious!

P.S. We are considering a step-by-step Deep Dive on vibe coding. If that interests you, use the rating below and leave us a comment.

That's it for today. See you in the next issue!

Reto & Fabian from AInauten