Ahoy, AInauts,

Welcome to a new issue of your favorite newsletter.

Today you get two prompts that do real work: one builds a website for you, while the other shows you what ChatGPT thinks it knows about you.

The reason was concrete: enrollment is open for the third, and possibly final, live round of our AI Employee Bootcamp. We used an AI employee to build a new test version of its landing page. The Bootcamp is designed for the German-speaking market and is conducted in German. We gave the agent the current URL, our brand rules, and a clear goal. Sixty-one minutes after the final go-ahead, the new site was live on its own domain. We cleaned the house while it worked. AI is still surprisingly unhelpful with that part 😁.

Below, we show you the complete workflow and the prompt. Then we audit ChatGPT's memory and explain why Kimi K3 matters even if you never open the model yourself.

Here is what we have for you:

Let's get started.

Every manager faces the same situations before lunch: a message to land, a meeting to run, a hiring call, a report due. The AI Report built 37 Claude prompts for exactly those moments, organised by the situations every manager faces. 

Copy the prompt, fill the brackets, run it in Claude, and get back 10+ hours a week. Oh, and it's free. 

All you have to do is subscribe to The AI Report, a 5-minute daily AI brief read by 400,000+ business leaders at IBM, AWS and Microsoft, and the full prompt pack lands in your welcome email. The newsletter and the prompts, both free. Subscribe and grab both

Subscribe and get 37 prompts free

Can you have a website built without knowing how to code? By now, the answer is yes.

Give ChatGPT Work/Codex or Claude an existing URL and explain what the new site should look like. The agent creates the content, design, and code. If you want, it can even publish the result on your own domain.

The catch: "Build me a beautiful website" often produces the usual AI-slop design. Purple gradient, rounded cards, an icon above every heading. A clear workflow and the right design skill produce a much better result.

The easiest starting point is a website you already own and want to redesign. The URL gives the agent your existing content, navigation, images, and trust elements. We prepared a prompt for exactly that. Copy, paste, and go.

Language note: the linked prompt document is in German - but your AI understands 😁.

Codex and Claude Code can both build strong frontends. We like to start by generating several design directions as images with GPT Image 2 or another image model. Once we choose a direction, the coding agent has something concrete to follow.

Visual note: some labels in the following design examples are in German.

Design skills help prevent the result from looking like every other AI-generated website.

Give your AI helper the necessary context, then let it build the site.

Pro tip: once the first version is ready, use the built-in Codex browser to click a heading, image, or button and leave feedback right next to it.

This saves a surprising amount of explanation. The agent immediately knows which element you mean, changes the relevant component, and checks the site again.

With ChatGPT Sites, the finished site can be published directly. Give Codex access to your domain and the agent can also handle the technical steps through the browser or Computer Use: determine DNS records, set the CNAME, and verify HTTPS.

Sites is rolling out gradually on paid plans, so it was not available in every account when we tested it. It appeared in our desktop app through a US VPN, and we were able to publish the page. That is a practical workaround, not a guarantee.

If Sites is missing, ask the agent for a static HTML version or production build. You can publish the folder through Netlify's drag-and-drop publisher. Its free plan supports a custom domain and SSL. Cloudflare Pages is another free option we use regularly.

The same workflow works with Claude. Claude Design helps with the visual direction, Claude Code or Cowork builds the site, and Netlify or Cloudflare handles hosting.

We tested this workflow with the existing landing page for our AI Employee Bootcamp, the prompt above, and our brand materials. The Bootcamp is aimed at participants in Germany, Austria, and Switzerland. Codex used GPT Image 2 to show us several design directions. We chose one, set a few constraints, and used the /goal command to define the final objective.

Then we left the computer running and cleaned the house 😁. We finished with a few short feedback rounds in the built-in browser. From the final go-ahead to a verified domain, the entire process took about 61 minutes.

Market and language note: the Bootcamp, its live sessions, community, materials, and landing page are all in German.

View the finished German-language Bootcamp landing page here.

Do not choose a complex platform with a store, login, and customer data for your first attempt. Start with a landing page, portfolio, or small company website.

Give the agent a clear outcome, your design rules, and permission to plan the technical steps on its own.

Instead of building page by page with AI, you let a digital employee deliver the finished website. Try it with one small, clearly bounded project.

AI will eliminate 300 million jobs in the next 5 years.

Yours doesn't have to be one of them.

Here's how to future-proof your career:

Start learning AI now

Our AI self-check from last week was popular with readers.

Everyone wants to know what ChatGPT can find about them online: photos, companies, old profiles, strange search results, the whole package.

That is the outside view. There is another useful question: What does ChatGPT already think it knows about you before you even search?

Facts. Relationships. Language. Projects. Writing style. Tools. Old goals. Half-remembered preferences. Things you mentioned once in passing that are now treated like permanent rules.

ChatGPT's memory is not memory in the human sense. It is carried-over working context. Some of it is gold. Some is outdated. Some was only a guess from the start.

Visual note: the labels in the following ChatGPT screenshots are in German.

We put together a prompt that lets you audit all of this in detail.

Language note: the linked prompt document is in German - but your AI understands 😁.

The most important instruction asks ChatGPT to distinguish clearly between facts, deductions, assumptions, and uncertainty. That is where useful context separates from fantasy.

If ChatGPT says, "You work quickly, take risks, and are driven by projects," that may be true. It may also be a pattern inferred from three launch chats.

If a point surprises you, ask:

Then you have three choices:

You can also manage this manually in ChatGPT under Settings and Personalization.

If you want something removed completely, deleting it from the memory summary may not be enough. Check old chats and files as well.

The public self-check shows you what other people can find about you with AI. This memory check shows you the context your own assistant is using.

If your assistant remembers you incorrectly, every result tilts slightly in the wrong direction. It is rarely dramatic. It is just annoying enough that you may eventually blame the model.

A few minutes of cleanup can improve every conversation that follows.

The new Kimi K3 model has the AI hype crowd on X excited. It builds spectacular interfaces, performs strongly in frontend coding, can be connected to Codex, and is being pitched as a replacement for expensive frontier models.

It is not that simple, although the hype is not coming from nowhere. The early benchmarks and hands-on reports look strong.

You will probably never open Kimi K3 yourself. Your ChatGPT history lives at OpenAI, your team works with Claude, and nobody rebuilds an entire AI workflow because one new model appears.

K3 can still influence what you pay for powerful AI and who gets to decide whether you can use it.

The Chinese company Moonshot AI released K3 a few days ago. You can try it through Kimi, Moonshot's API, or OpenRouter, although capacity is currently at its limit. OpenRouter currently forwards requests directly to Moonshot. It becomes operationally independent only when other providers host the model themselves.

Moonshot places K3 behind Claude Fable 5 and GPT-5.6 Sol, but its own tests already put the model in the top tier. The API costs $3 per million input tokens and $15 per million output tokens, about 70 percent less than Fable 5.

Cheaper tokens tell only half the story. In one coding comparison, K3 used roughly twice as many tokens and ended up costing about as much per task as GPT-5.6 Sol.

Kimi still puts pressure on US providers. We have seen the pattern before.

In early 2025, DeepSeek R1 showed how close an inexpensive Chinese model could get to OpenAI. A few days later, Nvidia lost more than $500 billion in market value in a single day.

Moonshot reports that K3 scales 2.5 times more efficiently than K2. Only a small part of the huge model is active for each answer, saving computing power. US chip restrictions increase pressure on Chinese labs to get more from less hardware.

Distillation is another part of the race. A model learns from the answers of a stronger model, and all major labs use versions of this technique.

Anthropic accuses DeepSeek, Moonshot, and MiniMax of collecting more than 16 million Claude responses through fake accounts. Technical progress, efficiency pressure, and learning from competitors are probably all contributing.

K3 is expected to become open weight. Moonshot plans to release the model files by July 27, allowing other providers to run K3 on their own servers instead of sending every request to Moonshot.

Kimi is not alone. Alibaba announced Qwen3.8 with 2.4 trillion parameters as another open-weight model and places it directly behind Fable 5 in its own tests.

These models will not run on your laptop. They are too large and require data-center hardware. Most users will still access them through a cloud or API provider.

The provider could eventually be based in Europe. A competitor does not have to beat ChatGPT at everything. Good enough, cheaper, and independent is sufficient for many tasks.

In June, the US government forced Fable 5 and Mythos 5 offline worldwide. We previously covered the Fable hype and then the dependency shock.

Fable returned with additional safeguards and is now included with Max and Team Premium. Pro and Team Standard users can access it through usage credits and receive a one-time $100 credit.

The same dependency appears with Palantir. A US provider controls technology that European authorities rely on.

France wants to replace Palantir at its domestic intelligence service, but the change is expected to take one to three years. What happens when a foreign partner changes the price, rules, or access?

Europe does not need to ban US models or send sensitive data to China.

It needs a credible second path: strong open-weight models, European data centers, and workflows that are not tied to a single provider.

K3 is less a ChatGPT replacement than a shift in bargaining power. It shows that frontier performance is getting cheaper and no longer comes only from California.

The question for you is simple: could you switch if you had to?

That is it for today. Thanks for reading, and see you in the next issue.

Reto & Fabian from AInauten