Good morning, AInauts,

Welcome to a new edition of your favorite newsletter.

This week, once again, there is less breaking news and more to read, watch, think about, and, of course, put into practice yourself.

Here is what we have for you today:

Let's go!

A PM can summarize research, draft a PRD, and mock up a prototype before lunch. The hard part starts when the team has to decide what actually gets built.

Jira Product Discovery gives product teams one place to capture insights, prioritize ideas with consistent frameworks, and build living roadmaps stakeholders can rally around.

And because it’s connected to Jira, the context behind every decision stays with the work—so developers and their agents know not just what to build, but why.

AI helps PMs move faster. Jira Product Discovery helps the whole team build with confidence.

Get Started for Free.

This issue is brought to you by:

Let's start with a slightly philosophical topic that still has a major impact on the results we get from AI every day.

If you have been reading us for a while, you know we are big fans of Shopify founder Tobi Lütke. Whenever a new interview with him appears, we are all in.

That happened again this week, and the conversation fits perfectly with something we have been thinking about for a while. The title: “Most Businesses Are Using AI Wrong.”

One idea from it that we want to explore today:

Being lazy at work used to mean producing too little. With AI, being lazy can now mean producing too much.

Online and inside Shopify, they call these “slop grenades”: low-effort AI output that colleagues toss back and forth.

Here is the exact clip from the interview:

A slop grenade is what happens when you let AI produce something and forward it without reading it. You add no value yourself. You do not even look at it.

A classic example: You tell AI to “make something,” it generates a wall of text, you say “looks good,” and send it to colleagues, partners, or customers.

Or you receive a long email that might be important. You read it, realize it is AI-generated fluff, and let your own AI answer it with even more fluff.

Sure, everyone saves time and looks productive. The people who pay the price are those who carry responsibility and have to check everything.

We see this everywhere right now: with clients, in our community, and in our own work.

AI produces, we delegate, and output piles up in project folders.

Reports upon reports. Summaries of summaries. Not all of it is useful.

Quite the opposite: Errors often creep in without anyone really checking them. Tomorrow's AI then works from yesterday's AI mistakes.

Eventually, nobody has a clear overview because nobody is doing enough of the thinking.

The results get worse and worse, and we do not know why.

That is exactly what happened in two second-brain setups we built. Too much went in, too much was produced, errors accumulated, and eventually the quality of the output dropped.

We depend on these systems, so we started fixing them.

First, we cut our own instruction file from 240 lines to 58.

Then we streamlined the entire folder structure.

Projects now live outside the brain. The brain only knows where to find them. Meeting recordings are processed and categorized, but stored elsewhere.

And the AI received new rules: produce less and be clearer.

And the most important rule of all: Think for yourself as much as possible. Stay creative. Do not simply accept an AI-generated plan and keep clicking.

Stay in the driver's seat.

It keeps your brain sharp, too.

In July, we wrote this: Read AI-written text anyway. An AI-like style says nothing about its quality. That still holds.

But the other side matters, too: Read it, yes. Forward unread AI output, no.

Let's respect the time of our colleagues and partners instead of bombarding them with fluff.

One more line from the interview that we fully agree with: Machines cannot take responsibility.

Do not use AI to make the judgment. Use it to build the environment in which you can make the judgment.

AI can produce. The direction it takes and what ultimately goes out are your responsibility.

You don't have to scroll every AI thread, track every new tool, or watch every demo. 

The Rundown AI breaks it all down for you — the latest AI news, tools, and tutorials in one free 5-minute email every morning. 

Trusted by 2M+ professionals at Apple, Google, and NASA.

Sign up to start learning.

With our partner HubSpot

Now let's turn to a topic specifically for companies.

UNBOUND is taking place in Boston this week. You may still know it by its previous name, INBOUND.

We have covered it for three years, not only because HubSpot is one of our partners, but because the event consistently features strong marketing strategies and practical AI experience.

Why is it called UNBOUND instead of INBOUND now? Inbound marketing was a method HubSpot more or less invented, and it became extremely popular over the past several years.

You can watch the spotlight keynote with the latest announcements here:

We have now spent several years working with AI, and naturally a lot has changed during that time. Growth, whether in customers or revenue, no longer fits neatly into one department because AI now cuts across marketing, sales, and service.

The question is no longer whether your company should use AI. The question is how to use AI effectively and, above all, safely.

One topic is especially important for companies right now: context.

AI without context produces output without value. For many of us, that has long been obvious thanks to second brains, skills, memory, and files over tools.

Hopefully, you are already using some of those ideas.

For a company, however, the problem is harder. Which data should be available to which AI? Which data should be shared at all, and which should stay private?

HubSpot is building the same kind of context layer for CRM.

The models are the same. The context is not.

HubSpot co-founder Dharmesh Shah built his keynote around exactly that idea: the same AI can produce completely different results depending on its context.

For companies, it is essential that AI always has access to the latest and most relevant information about the business, even though that is often difficult to implement.

Only use tools with interfaces

For this to work, the AI tool you choose, whether Claude, ChatGPT, Langdock, or something else, must be able to access the other tools where your data lives.

That is why we have long favored tools that offer MCPs or at least a strong API.

Give employees a safe framework

Safe adoption may be the most important issue right now. Employees need to be empowered. But if everyone can install Claude Code and arbitrary packages, serious problems can emerge quickly.

One possible safeguard: We often let our AIs work inside systems and tools whose architecture limits the number of critical mistakes they can make.

Of course, problems can still occur even in the best tools. But this at least creates a buffer.

Tools such as HubSpot naturally cover these points, which is why they are useful for many companies. Still, we think the principle applies more broadly.

In August, the question was why gains from AI efficiency were not showing up in the numbers. HubSpot's answer: Because AI sits next to the data instead of directly on top of it.

Language note: HubSpot's live webinar and the linked registration page are in German.

On September 30, HubSpot will show what it looks like when agents work with real context in a free live webinar.

Two scenarios, demonstrated live:

The details:

📅 Wednesday, September 30, 2026, at 10:00 a.m.
🎬 A real live demo in Marketing Hub, Sales Hub, and Agent Hub, presented in German
📼 If you cannot attend live, every registered participant receives the recording

If you work in marketing or sales, lead a team, or simply want to see what agents can do with real context, take a look. It is relevant whether or not you use HubSpot.

It is always useful to see how other teams work.

To wrap up, here is something practical.

Two weeks ago, we shared the first two research-backed prompting tips: Do not end with “right?” and use a role instead of asking for “step-by-step” reasoning.

New models change old prompting techniques. So today, here are two more ideas you can apply immediately.

Meta tested 15 models, including GPT-5.5, Claude 4.7 Opus, Gemini 3.1 Pro, and twelve others, using prompts that contained between one and twelve rules at once.

Rules such as: Exactly three paragraphs. Fewer than 150 words. No emojis.

With eight rules, a model follows each individual rule only 41 percent of the time. All eight together: 5.7 percent.

Twelve of the fifteen models already fall below 50 percent when the prompt contains just three rules.

One detail surprised us, too:

Numeric rules, such as word or paragraph counts, cause far more problems than word-based rules, such as required or forbidden terms.

“Under 150 words” is therefore not a particularly reliable instruction.

Bad:

Better, first pass:

Second pass:

To be fair, the review pass recovers only one or two rules according to the paper, not all of them. Still, it makes sense to work through a long list of requirements in stages.

→ Paper: Large Language Models Can Follow Instructions, But Not Many at Once

For years, we preached: Give AI examples. Few-shot prompting. Show it what the result should look like.

EPFL, Apple, and Mistral tested this on math problems. With eight solved examples in the prompt, the model answered 74 percent correctly. Without examples: 84 percent.

The paper's explanation: Fixed examples push the model to copy style and format instead of thinking.

Today's models are far more capable and can reason well. In other words, the stronger the model, the more examples may get in its way.

Important: This applies only to top models such as Opus, Fable, GPT-5.6, GPT-6, and similar systems. One weaker model in the test actually improved with examples.

Bad:

Better:

No example. Instead, a clear goal and real context.

One caveat about the study: It used small 7B models and tested only math problems.

Even so, the findings seem plausible to us. We see the same pattern in our daily work: Give strong models room to work things out for themselves, guided by a clear goal.

→ Paper: Soft Guidance Starts to Outperform CoT Prompting as LLMs Improve

Yes, elaborate prompting is not quite as important as it used to be. But you can still get much better results when you understand a few rules and quirks.

At the same time, prompting remains a field that changes a little with every new model.

Good for us. It means we will never run out of content and things to test. 😉

That is it for today. An issue about producing less that somehow became too long again. We are working on it 😁.

As always, thank you so much!

See you next week.

Reto & Fabian from AInauten