Daniel Brugger

Law/AI, Head Legal Tech/AI ( Senior Legal AI Advisor ( Founder and Podcaster

[Daniel Brugger @_DBrugger]

Danke für die Initiative! In der Schweiz haben wir genau das gleiche Problem.

[Daniel Brugger @_DBrugger]

The Pre-Mortem Prompt: How to Break AI Sycophancy

Six months from now, your project has failed. Your LLM knows why. You just have to ask it the right way.

Anyone who works with Claude or another AI tool knows the pattern. You share an idea, and the model responds favorably. This phenomenon is commonly referred to as sycophancy. The pre-mortem prompt is a simple and effective antidote. But before we get to it, it's worth looking at where the idea comes from.

❓ Reversing the question
In his research on decision-making, the psychologist Gary Klein observed a recurring pattern. Teams would commit to decisions that individual members privately considered problematic, yet those concerns never made it to the table. The reason was group dynamics. Anyone who points out risks during the optimistic mood of a project launch gets cast as a naysayer. With the pre-mortem, Klein flips the timeline (https://t.co/YM1YBuiKZD). Instead of asking "what could go wrong", you start from an assumption: the project has already failed. The task is now to reconstruct what killed it.

Klein himself describes the technique as follows: "A Pre-mortem is an exercise that typically begins after a team has been briefed on a plan. The team leader will ask his or her team to imagine that the proposed plan has failed. As a team, everyone contributes to an exercise where potential threats and hurdles to the plan are generated. The goal of the Pre-mortem Method of Risk Assessment is to increase plan success rate." (https://t.co/DwJi1MGIvF). This shift changes the social dynamic. No one has to play the skeptic anymore, because failure is already a given.

🤖 Why the method is especially useful with LLMs
The dynamic Klein observed in teams is even more pronounced when you're talking to a language model. An LLM is not just conflict-averse, it is systematically trained to please the user. The pre-mortem prompt forces the model into a different role, one in which it has to name the reasons a project might fail.

Building on this idea, @itsolelehmann recently built a sophisticated pre-mortem skill (https://t.co/bXjQaUjWP6). Both the skill and his article on it come highly recommended. As a starting point, a simpler prompt does the job in most cases. Something like this: "Imagine we are six months down the road. The project I just described to you has failed catastrophically. Walk me through what happened. Name the most likely causes, identify the underlying false assumption, and describe the early warning signs I should have seen along the way."

You can apply this to almost any decision, from a strategic hire to a product launch to the rollout of new software. The real appeal of the pre-mortem prompt is that the model takes on exactly the role no one on a team likes to play voluntarily.

The LLM has no reputation to protect and no relationships to spare. It names the risks that would otherwise go unspoken, and it does so without the social cost a human dissenter might pay. Anyone who runs this prompt before an important decision brings to the table exactly the voice Klein found missing in his work.

____
More on AI and Law: https://t.co/knNpbf80R4

[GREG ISENBERG @gregisenberg]

https://t.co/zB1WgXzFAz

[Daniel Brugger @_DBrugger]

https://t.co/N4WOMHu9VU

[Daniel Brugger @_DBrugger]

Identity Before Technology: What We Can Learn from Atomic Habits for the AI Transformation

In "Atomic Habits" James Clear presents a model that describes three layers of behavior change. Outcomes sit on the outside (what you get). Processes come next (what you do). And at the core lies identity (what you believe). Most people start their behavior change from the outside in. They focus on what they want to achieve and then try to find the right process to get there (outcome→process→identity). Clear argues that the reverse direction works better. You start with identity. Instead of focusing on what you want to achieve, you focus on who you want to become. The habits follow naturally (identity→process→outcome). I believe this approach can also be applied to the AI transformation in the legal market.

The Outcome Trap
Most AI initiatives in the legal sector begin with the same pitch: faster work, lower costs, more consistency, and so on. The language is purely outcome-oriented. That is not wrong, but it is superficial and can cause the transformation to fail: an expensive legal tech tool, a few enthusiasts, but no organization-wide transformation.

The Process Plateau
More advanced organizations move to the second layer, the process level. They redesign their workflows, clean up their document management, and build prompt libraries or "workflows" in legal tech tools. They start thinking about which processes AI can actually support. That is better. But it can stall too. You can redesign a process as many times as you want. If the people executing it do not believe in what it stands for, they will not use it.

The Real Starting Point
Organizations that will lead the AI transformation should, in my view, ask themselves a different question. The right question is not "How can AI make us faster or better?" or "How should we restructure our workflows?" but rather "What kind of firm or organization are we?" That is the identity level.

Identity determines which processes feel natural and which feel forced. It decides whether a new AI tool is adopted or abandoned after the first disappointing result. It is the difference between a law firm where partners say "We are an AI-forward firm" and one where they say "We tried that AI thing."

Identity First, Then Everything Else
For law firms, this means the AI conversation must start at the partner retreat, not in the IT department. It must begin with beliefs about what legal work means in 2026 and beyond and where the firm's distinctive value lies. Once that is clear, you can move on to designing processes. The outcomes will follow (identity→process→outcome).

But anyone who skips the identity level will keep buying tools that never see widespread adoption and building workflows that feel unnatural. Better start with identity.

___
Originally published as a German-language blog post on https://t.co/PE2OQJRwPJ: https://t.co/u8x0k8L75y

[Marc Andreessen 🇺🇸 @pmarca]

My information consumption is now 1/4 X, 1/4 podcast interviews of the smartest practitioners, 1/4 talking to the leading AI models, and 1/4 reading old books. The opportunity cost of anything else is far too high, and rising daily.

[Daniel Brugger @_DBrugger]

@zackbshapiro Great article. Thanks for sharing

[Daniel Brugger @_DBrugger]

https://t.co/VNeoXYyziS

[Daniel Brugger @_DBrugger]

https://t.co/CZ30ayiH8m

[Daniel Brugger @_DBrugger]

https://t.co/A7H348dyYh

[Daniel Brugger @_DBrugger]

🚀 Join https://t.co/knNpbf80R4 - Where Law Meets Digital Transformation

IusBubble is a digital community for lawyers and everyone interested in the future of law and its digital transformation.

What started as an idea is now a growing community of over 780 members exchanging ideas in English, German, and French.

IusBubble hosts regular webinars focused on practical, real-world use cases. Just yesterday, members joined a step-by-step tutorial on how to connect their AI tools to an MCP server. All sessions are recorded and available for members.

👉 Join over 780 members and start bubbling ideas into reality.

Registration form: https://t.co/lh85aent5H

[Daniel Brugger @_DBrugger]

Learn to Think Before You Start to Prompt

In a concise blog post (https://t.co/iP5jrXEyfY), Reto Gubelmann highlights a crucial aspect that frequently gets overlooked:

"At a university, efficiency is not the goal. Formation of habits, virtues, and acquisition of deep skills is. You acquire those by doing the nitty-gritty work of a BA student yourself. Not because you will do these tasks without GenAI when on the job (there, it is about efficiency), but because it is the only way to acquire the grit, depth, precision, and flexibility of thought that will allow you to succeed in a labor market shot through with AI."

A thought-provoking blog post on the use of AI in legal education.

[Daniel Brugger @_DBrugger]

KI-Assistent für Schweizer Rechtsprechung in Deinem KI-Tool einrichten (Schritt für Schritt Tutorial)

Mit der neuen MCP-Integration von Lexi Search kann man jedes kompatible KI-Tool direkt mit LexiSearch verbinden (https://t.co/a8VI85OlQv). Dadurch kennt Dein KI-Tool aktuelle Schweizer Gerichtsentscheide und kann sie zitieren. Wie das geht, zeigt Lionel Voser im https://t.co/knNpbf80R4 Webinar Nr. 28 vom nächsten Dienstag, 27. Januar 2026, 12.00-12.30h.

Im Webinar erwartet Dich ein Schritt-für-Schritt-Tutorial zur Verbindung Deines Assistenten mit Lexi Search per Model Context Protocol (MCP). Gezeigt wird es am Beispiel von Claude und Mistral (weitere Infos in der Bubble). Am Ende des Webinars hast Du einen fertig eingerichteten Assistenten, der im Chat aktuelle Schweizer Rechtsprechung durchsucht und Entscheide sauber zitiert.

💡 Warum ich die Lösung von LexiSearch spannend finde: Durch das Model Context Protocol (MCP) werden Rechtsdaten - hier Schweizer Rechtsprechung - zu einer andockbaren Infrastruktur. Anstatt auf teure Speziallösungen oder zusätzliche Legal-Tech-Plattformen angewiesen zu sein, können Nutzer ihr KI-Tool frei wählen und dieses direkt mit einer Datenquelle verbinden. Dadurch entsteht kein Lock-in, sondern ein offenes, modulares System.

---

https://t.co/knNpbf80R4 ist eine digitale Community für Juristen sowie für alle anderen, die sich für das Recht und seine Digitalisierung interessieren. Die Mitgliedschaft ist kostenlos. Die Registrierung ist einfach, aber kontrolliert, um Bots und störendes Verhalten in der Gemeinschaft zu vermeiden. Anmeldung auf: https://t.co/yIofX22ZwZ

[Joël Niklaus @joelniklaus]

Super excited to be heading to Davos next week for my first @wef!

I'll be participating in a roundtable discussion on legal AI in Davos on January 20th. Thanks to @RobertMahari for the invitation and organization! I'm looking forward to the conversation with @_DBrugger, Paul Chow, Dr. Rehana Harasgama-Zehnder, Ilona Logvinova, Philippe Gillieron, @alex_pentland, and Roland Vogl.

The discussion will focus on how legal AI is evolving beyond document automation toward genuine reasoning support. We'll explore how law firms and in-house teams can use AI strategically, how data-driven analysis is changing business models and client expectations, and what new questions around trust, liability, and professional responsibility are emerging.

Leading Beyond Boundaries brings together researchers, business leaders, and policymakers in Davos to discuss AI, justice, healthcare, and other areas shaping our future. The initiative combines invite-only roundtables with public sessions for cross-disciplinary conversation.

[Daniel Brugger @_DBrugger]

🚀 https://t.co/EPwmDNUGtR: Gebaut, getestet, und nun unter Live-Bedingungen bestanden.

Man kann vieles intern vor dem Launch testen. Aber der wahre Test kommt erst unter echten Bedingungen. Genau das konnte ich kürzlich mit Marco Candinas durchspielen:

1️⃣ Marco hat seinen Onlinekurs „Praktische Anwendung von Künstlicher Intelligenz im juristischen Alltag“ auf Lawcad selbstständig mit dem einfach verständlichen Drag-and-Drop-Editor erstellt.
2️⃣ Der Kurs wurde erfolgreich an mehrere Personen verkauft.
3️⃣ Marco hat über die Plattform seine Kursbeteiligung (95 % des Kurspreises) angefordert und ich habe ihm diese ausbezahlt.
🔁 Kreis geschlossen.

Was intern gut aussah, hat sich im Live-Betrieb bewährt: Erstellung, Verkauf, Abrechnung, Auszahlung. Alles funktioniert.

Danke Marco fürs Vertrauen und fürs Mitgehen als erster echter End-to-End-Test. 💪

[Daniel Brugger @_DBrugger]

https://t.co/qJcG5PBj9e

[Daniel Brugger @_DBrugger]

France’s First Documented AI Hallucination Case

On the website “AI Hallucination Cases Database” (https://t.co/TV7K77JEhL), the first hallucination case from France has been published (available here: https://t.co/hGNejntz0d)

[Daniel Brugger @_DBrugger]

Jagged intelligence (https://t.co/zpdDRuHw9A) means that AI models are not uniformly intelligent. They perform certain tasks extremely well, while failing catastrophically at others.

Unfortunately, it is by no means always obvious what large language models can do reliably and where dangerous failures may occur. With growing experience, one can develop a certain intuition about the situations in which a model is robust and where caution is required. However, this intuition does not emerge automatically; it is built through observation, testing, trial and error, and deliberate reflection on the behavior of the tools. This is why it is important to actively experiment with these tools in order to develop a feel for their strengths and weaknesses.

[Andrej Karpathy @karpathy]

Jagged Intelligence

The word I came up with to describe the (strange, unintuitive) fact that state of the art LLMs can both perform extremely impressive tasks (e.g. solve complex math problems) while simultaneously struggle with some very dumb problems.

E.g. example from two days ago - which number is bigger, 9.11 or 9.9? Wrong.
https://t.co/dUrR6wm8GC

or failing to play tic-tac-toe: making non-sensical decisions:
https://t.co/XarwfUBtod

or another common example, failing to count, e.g. the number of times the letter "r" occurs in the word "barrier", ChatGPT-4o claims it's 2:
https://t.co/xpffK2r0pv

The same is true in other modalities. State of the art LLMs can reasonably identify thousands of species of dogs or flowers, but e.g. can't tell if two circles overlap:
https://t.co/HCXxBxosAu

Jagged Intelligence. Some things work extremely well (by human standards) while some things fail catastrophically (again by human standards), and it's not always obvious which is which, though you can develop a bit of intuition over time. Different from humans, where a lot of knowledge and problem solving capabilities are all highly correlated and improve linearly all together, from birth to adulthood.

Personally I think these are not fundamental issues. They demand more work across the stack, including not just scaling. The big one I think is the present lack of "cognitive self-knowledge", which requires more sophisticated approaches in model post-training instead of the naive "imitate human labelers and make it big" solutions that have mostly gotten us this far. For an example of what I'm talking about, see Llama 3.1 paper section on mitigating hallucinations:
https://t.co/pjuxoIOJCY

For now, this is something to be aware of, especially in production settings. Use LLMs for the tasks they are good at but be on a lookout for jagged edges, and keep a human in the loop.

[Daniel Brugger @_DBrugger]

Plato, Writing, and Our AI Debate: An Old Pattern Repeats Itself

[AI] is “inhuman” because it pretends to offer knowledge that is, in truth, no knowledge at all. True knowledge cannot exist in an external medium. [AI] is a manufactured product, static and incapable of replacing real thinking. Those who use [AI] weaken their memory, for they rely on an external support instead of training their own judgment. [AI] is a technology that produces an outward appearance of wisdom while letting inner thinking wither.

Sounds like today’s AI debate, doesn’t it? Yet these concerns are far older. They come from Plato, not directed at artificial intelligence, of course, but at the technology of writing. I merely replaced “text” or “writing” with “AI.”

Discussions about ChatGPT, AI assistants, and digital tools feel new, but the underlying questions are surprisingly old. Janique Brüning notes in her recent article (KI als Herausforderung für das juristische Studium, ZDRW 4, 2024, pp. 291 ff., 299 https://t.co/g5sZ6ySeR7) that Plato, in the dialogue Phaedrus, raised almost exactly the same objections that many people today express about AI or other digital tools: writing, he argued, was a dangerous technology. It was “inhuman” because it simulated knowledge that was not real knowledge. True knowledge, Plato claimed, could exist only in the mind, not in any external medium. Writing, as a manufactured object, was static, non-dialogical, and incapable of replacing genuine thought. Those who relied on written texts would weaken their memory, depending on an external support instead of cultivating their own judgment. Writing lacked interactivity: you can question a human being, but not a text. A text merely repeats the same words, whether they fit or not. It is a technology that produces outward pseudo-wisdom while allowing inner thought to atrophy.

Reading the objections that Brüning recounts, one hears today’s AI debate: “AI makes us dumber,” “AI produces superficial knowledge,” “AI destroys our ability to think for ourselves.” The pattern is timeless. Every new technology is first perceived as a threat to thinking before it becomes a natural tool for thinking.

Perhaps we stand at exactly this threshold with AI today. The question is not whether AI will change our thinking. It will, just as every media innovation since writing has done. The question is how we integrate AI in a way that improves our learning and our work.

IusBubble Blogpost: https://t.co/oQfCRLQndo