Ilya Boyandin

Researchers at ETH Zurich just published a paper saying context files like CLAUDE.md reduce agent task success rates and cost 20% more to run. I have a 500-line CLAUDE.md. My orchestration system references it across 38 automated daily jobs. So either I've been doing this wrong for months, or something else is going on. Here's what the research actually found. Across most coding tasks, Claude already knows how to traverse a codebase. Adding a context file gives it more to process for tasks where that context is irrelevant -- noise dressed up as instructions. Standard code projects with existing documentation? The CLAUDE.md adds drag, not signal. The exception they identified: repositories with no code architecture to traverse. Wide, flat, document-heavy systems with highly personal conventions the agent can't infer from the files themselves. That's the system I run. My CLAUDE.md isn't a coding guide. It's operating context for a personal intelligence system. Project handoff formats. Social routing rules across three platforms. Voice conventions. Telegram bot IDs. None of that lives in any traversable file. The research is right. For standard coding projects, less is more. But if you're running a personal AI system that operates on your own data and workflows, the rules change. Your conventions aren't in the code. The CLAUDE.md isn't documentation. It's the piece the agent can't discover on its own. Know which you're building.

2

Want to learn data science, machine learning, and AI and happen to live in or around Zurich? Apply to ETH Zürich's DAS (Diploma in Advanced Studies) Data Science! I recently completed the DAS program after two years of part-time study. The DAS allows you to choose from a range of master's-level classes across data science and adjacent disciplines (computer science, mathematics, statistics, neuroscience, and electrical engineering). There are over forty classes to choose from. I managed to squeeze in eight classes over the two years, all of which were uniformly excellent. The teaching was among the best I have experienced, and the courses were incredibly up-to-date, with us often studying cutting edge material that was only released in the preceding weeks and months. The assignments were also well thought-out and instructive. Going into this program, I already had a solid foundation in machine learning, but the diploma allowed me to fill in many gaps and to spend much more time and effort investing in areas I was less familiar with: reinforcement learning, Bayesian machine learning, and more esoteric neuroscience-based approaches. Even in the areas that I was already familiar with, there was plenty of learning to be had from large language model watermarking, fundamentals of natural language processing theory, and a more statistically rigorous view of model ensembling, to give a few examples. Another awesome benefit of the program is the networking. There are other part-time professional students like myself, on top of the usual ETH master's and PhD students, so I got to network with, collaborate with, and become friends with many awesome professionals and students. My only criticism of the program is that I would happily do even more. I would eagerly do another few years and earn another set of ECTS credits to get an MAS (Master of Advanced Studies) if it were offered. Hint, hint, ETH Zurich | School for Continuing Education ;) I have done a lot of part-time study over the years and I can state that this DAS was my favorite so far. Highly recommended for those with any interest in the field! Diploma information: https://lnkd.in/eDKwtMCZ Class list: https://lnkd.in/ehutbZ-g

210

13 Comments

🚀 ETH spin-off Invariant Labs acquired by cybersecurity leader Snyk! Founded in 2024 by computer science professors Martin Vechev and Florian Tramèr and three of their graduates, Marc Fischer (CEO), Dr Mislav Balunović (former CTO) and Luca Beurer-Kellner (Chief Scientist), Invariant Labs has made headlines with its #AI security innovations – now joining forces with Snyk to tackle next-gen threats. Congratulations to the whole team! 👏 From groundbreaking research to real-world impact: another success story from the ETH Zürich ecosystem. Read more here: https://lnkd.in/dJ2sSpku ETH Zürich ETH AI Center #DeepTech #CyberSecurity #SpinOff #AIsecurity

47

1 Comment

Just kicked off a new paper serialization: FallacyTag — on surfacing fallacies and reasoning structure in the wild. Two posts are live: 1. Why I’m Writing About Fallacies in Public 2. The Architecture of Avoidance (The problem statement — and a whiff of solution.) Feedback welcome. ToC: https://lnkd.in/ge-H3Dcy

18

1 Comment

Delete your CLAUDE.md files" That appears to be a highly discussed topic within the AI agents community at present. There have been over 1000 posts on X today engaging in a debate about it. It seems that a significant portion of these discussions may not be fully addressing the complete scope of the study. ETH Zurich has recently published new research investigating whether CLAUDE or AGENTS MD files are beneficial for coding agents. 138 real GitHub issues. 4 coding agents. one question: do context files help agents ship better code? the success metric was simple: did the agent produce a working patch? not tokens. not speed. working code. two findings came out of it. most people are only sharing one. auto-generated context files (the ones from running /init) hurt task resolution by 0.5-2% and bump inference cost by 20%. that part is true. but human-written ones improved performance by 4%. same study. different number. the reason auto-generated files hurt is kind of funny once you see it. agents are obedient. they follow whatever's in the file. so when a generated file adds unnecessary requirements, the agent does more testing, more file traversal, burns up to 22% more reasoning tokens, and still produces worse results than if you'd given it nothing. now here's where it gets interesting. when researchers stripped all existing documentation from the repos, auto-generated files went from hurting to helping by 2.7%. repos with good documentation don't need a generated context file. the agent was already going to find that information on its own. /init just repackages what's already there. what actually works is telling the agent things it cannot figure out from your codebase itself. which test runner to use. which package manager you prefer. non-obvious repo-specific commands. the researchers tracked actual tool usage to prove this. mention 'uv in your context file and the agent uses it consistently. don't mention it and the agent almost never reaches for it on its own. so no, don't delete it. if you ran /init and committed whatever came out, that's the thing to delete. write a few lines yourself instead. the study doesn't say CLAUDE.md is useless. it says auto-generated, bloated ones are. just stop letting the loudest take win over the actual data

11

Marco Gorelli will deliver a keynote at PyData Berlin on September 2nd, speaking about "Narwhals: enabling universal dataframe support." As a core contributor to both pandas and Polars, Marco has a unique perspective on dataframe interoperability challenges. His talk will cover how libraries can work seamlessly across different dataframe backends and what this means for the broader data science ecosystem. For those working with multiple dataframe libraries or building tools that need to support various backends, this will be a valuable session. 🎟️ Register now: https://lnkd.in/gE2DZKgu  #PyDataBerlin #DataScience #Python #OpenSource

28

I'll provide a counter to this. RDF is an abstraction, which comes down to a few core principles: 1. Most identifiers for resources should be globally unique. This means that you can reference an entity even if it is not within your database. 2. All data should be cleanly decomposable into a labeled directed cyclic graph. 3. A schema should exist that identifies how properties act abstractly (e.g., cardinality). 4. Namespaces help make resource identifiers global. If you manage them, you can represent your data as RDF. That's it. The reason that all too much of what passes for data out there isn't RDF is because it fails to make identifiers unique, it conflates matching things with matching things, or it has no internal consistency with regard to how properties are expressed, putting too much of the model into the data. Give me a data table and I can express it as RDF, I just have to take that data out of its container (with all of its implicit assumptions) to do so.

16

9 Comments

Spreadsheets: "Just rows and columns." The Reality: - 17 empty columns for spacing - 3 unrelated tables - Notes in random cells Spreadsheets are optimized for human eyes, not for standard ingestion pipelines. We went deep on solving this, combining pragmatic heuristics with graph modeling to segment sheets into meaningful regions before retrieval. Matt Kauffman breaks down the engineering here 👇 https://lnkd.in/gFVQAcj3

2

Aligning a group around a single, coherent model for a complex domain is rarely straightforward. It often requires more than just technical skill; it needs deep, ongoing collaboration. A VirtualDDD session back in 2022 with Stefan Hofer and Henning Schwentner explored this, drawing lessons from the practical experience of co-authoring a book. The talk focused on how the sustained effort of writing together provided insights into collaborative modelling. The session highlighted that the iterative process of crafting a shared narrative, continuously negotiating meaning, and solidifying a common language directly informs how we approach modellingsoftware. It showed how real-world, long-term collaboration builds shared understanding. Watch the session: https://buff.ly/pz4vXX5 #DomainDrivenDesign #CollaborativeModeling #SoftwareDesign #SharedUnderstanding

8

Love what Fiona Leibundgut posted and totally agree with Katia's view on "do only what is necessary" when it comes to the use of AI. We have to be smarter about when to use AI and if its use doesn't make sense from a social, environmental, or even financial perspective... why are we doing it! Climate debt is real so why are our business cases not assessing environmental impact criteria and how our technology choices impact sustainability? Clearly we all have to start to question "Things that make us go hmmm!" #SustainableArchitecture #SustainableIT #ResponsibleAI #LCAW25 #ResponsibleAI

7