AI-HOY, AInauts!

Welcome to a new edition of your favorite newsletter.

As always, plenty is happening in AI Valley. Alibaba has launched another massive model, Qwen3.8-Max, and an OpenAI model has solved ten decades-old math problems.

For less than EUR 2,000 in compute costs.

Interesting. But it does not change much about our Thursday.

So today, we are back with several practical topics:

Let’s get started.

When you type a prompt, you summarize. When you speak one, you explain. Wispr Flow captures your full reasoning — constraints, edge cases, examples, tone — and turns it into clean, structured text you paste into ChatGPT, Claude, or any AI tool. The difference shows up immediately. More context in, fewer follow-ups out.

89% of messages sent with zero edits. Used by teams at OpenAI, Vercel, and Clay. Try Wispr Flow free — works on Mac, Windows, and iPhone.

Start flowing free

Let’s start with a topic we genuinely had a lot of fun with. (And one that taught us quite a bit about ourselves.)

We suspect many AInauts will have a similar experience.

We recently listened to The Random Show with Tim Ferriss and Kevin Rose. We love hearing the two of them exchange ideas. Then, while looking around X to see what else they were working on, one post caught our attention:

Once again, the idea is to use what our AI chatbots already know about us to produce insights about our lives.

Kevin Rose built a prompt for exactly that: the Reflection Engine.

It is a Markdown file with 22 questions. You attach it to a fresh chat, write one sentence, and the model searches through everything it knows about you: chat history, memory, and uploaded files.

What comes back is a fairly ruthless portrait. Blind spots, contradictions, your most expensive habits, and more.

First, we ran it against our chat history: three years of conversations.

Language note: The next four screenshots show our German source material. The English captions translate the relevant content.

Then we ran the same prompt against our Second Brain in Obsidian: Daily Notes, voice-memo transcripts, annual reviews, and more.

Same prompt. Same person. Two very different results.

We already use ChatGPT and similar tools to discuss plenty of personal topics. The memory these tools have built can be unsettling.

It is remarkable how convincingly they can assemble psychological profiles from the data points they already have.

Disclaimer: We are not therapists. Do not use AI as a therapist or treat its output as a diagnosis. It can be wrong. For serious concerns, speak to a qualified professional.

As you may know, we keep a Second Brain where we deliberately reflect and journal. We let the AI loose on that material too. The result was even more impressive, although sometimes painful. 😉

Here is one more question and answer. It is fairly personal, but we are among friends.

We easily spent two hours with the results. They were genuinely impressive.

The real lesson: the prompt is not what determines the result. The corpus does. If your AI only knows you from shortening emails, it does not really know you.

If you have a Second Brain, discuss personal topics with your AI, and deliberately feed it useful material, the insights can be excellent.

There was one more detail we liked: every answer includes a confidence score and must provide counterevidence plus concrete possible actions.

You can copy that pattern directly into your own prompts. Competitive analysis, candidate evaluation, strategy review. Each one becomes more useful immediately.

Use your personal account, not a company account. First, check whether memory and past-chat search are enabled in your settings. Without them, the exercise goes nowhere. Use the newest model and a high reasoning level.

Most importantly: this is not therapy or a diagnosis. The result is sensitive. Do not paste it into a team chat.

You can download the Markdown directly from GitHub, attach it to a chat, and write:

Please review the attached file and complete all of its tasks.

We are honestly torn. This kind of exercise can quickly drift into horoscope territory.

We also sometimes worry that we will regret being this transparent with AI. In fact, we are almost certain that we will.

Still, it was so useful for us that we wanted to share it. With the right underlying data, it is excellent.

We identified at least five changes we want to make and genuinely believe will improve our lives.

P.S. This is exactly why we have been talking about Second Brains for months. Not because folders are exciting, but because what you collect eventually determines how well your AI can work.

You're spending 40 hours a week writing code that AI could do in 10.

While you're grinding through pull requests, 200k+ engineers at OpenAI, Google & Meta are using AI to ship faster.

How?

The Code newsletter teaches them exactly which AI tools to use and how to use them.

Here's what you get:

Sign up and get access to the Ultimate Claude code guide to ship 5X faster.

Join 200K+ engineers

A quick change of topic, but another practical issue that has kept us busy over the past few weeks.

You may have noticed it too: several genuinely strong new models arrived recently, from Anthropic and OpenAI. Yet some of our skills and workflows became worse instead of better.

Language note: The next two screenshots show German user feedback. Their meaning is translated in the captions.

It took us a while to understand why. Now we have it under control, and we want to share the lessons as usual.

Every new model behaves differently from the one before it. Ideally, it is smarter. But existing workflows can still break.

Imagine you trained someone two years ago. They were inexperienced, so you documented everything: check your work again, think step by step, here is the folder structure.

Today, the same person is very good at the job. Yet every morning, you still hand them the same note with the same detailed checklist.

The note now slows them down more than it helps.

That is exactly what happens with your prompts, skills, and routines.

Anthropic documented this itself and removed more than 80 percent of the Claude Code system prompt without a measurable drop in performance.

Their own conclusion was that they had overconstrained Claude Code.

The new models are substantially more capable than the previous generation. They already do many things on their own and need fewer instructions.

If you still give them those instructions, they will try to follow them. That is where problems appear.

The migration guide is even more explicit: if your prompt still says “check your work,” the model may now check it twice.

That does not make the result better. It only makes it slower and more expensive, and sometimes, oddly enough, worse.

We audited the CLAUDE.md file for our Second Brain, the central instruction file for the entire system. It had become bloated with micro-instructions.

Since Opus 5, we had been less happy with the output, so we cleaned it up.

In the end, 240 lines became 58.

Language note: The table below is in German. It shows the cleanup from 240 to 58 lines, a 76 percent reduction.

Here is what to check in your prompts, skills, and other instructions if you want them to perform better.

You can usually remove:

What should remain is everything the model cannot infer on its own: pitfalls, conventions with reasons, and your tone.

We published the complete cleanup prompt and the long version in Friday’s Deep Dive. Language note: this linked Deep Dive is currently available in German.

The real point applies to more than cleanup projects.

We are used to tools staying the same while we get better. With AI, the reverse is true: the tool improves while you stay the same.

Whenever a new model arrives, watch how it behaves before you keep running all your old prompts unchanged.

This takes half an hour, and it is absolutely worth that half hour.

We expect to clean things out around every six months. We also expect to rebuild many of our skills every few months. That is fine.

AI helps us do it and keeps getting better at the job.

To close, here is a good story for all the future parents among the AInauts.

The founder of HeyGen went on parental leave.

Instead of leaving inquiries unanswered, he cloned himself. An avatar with his face and voice handled the initial conversations. An agent in the background read internal systems, took notes, coordinated with colleagues, and more.

Of course, HeyGen is an AI avatar company, so we cannot rule out a certain marketing-stunt quality.

OpenClaw also ran in the background as the agent.

Every prospect knew they were speaking with the AI version.

After eight weeks: 2,741 prospects with at least one live call, 132 paying customers, 37 enterprise opportunities. And a baby.

In experiments where AI gets a lot of freedom to act, we often find the failures more interesting than the successes.

This one had failures too:

The good part: the agent learned from these mistakes and wrote them down as rules for the future.

The HeyGen and OpenClaw setup is also genuinely simple and can be rebuilt by almost anyone.

The founder even points to the Obsidian plus Claude Code approach, the simpler model we use too.

For anyone who wants to go deeper, the code is available as the MIT-licensed liveavatar-sales-agent on GitHub (we have not tested it ourselves). You can also try the agent live.

Posts like this always contain plenty of marketing and create the impression that an agent can now do everything alone.

And yes, OpenClaw, Codex, Claude Code, and similar tools can complete a remarkable amount of work autonomously.

Autonomous, not unsupervised.

Not everything runs autonomously for us either. We still want to understand what is happening and approve important actions.

The idea of AI clones and parental leave without email is still appealing, and probably not that far away.

That’s it for today. As always, thank you very much for reading.

And if you actually run the Reflection Engine, we would love to hear whether it hurts for you too.

See you this weekend!

Reto & Fabian from AInauten