Why AI Setups Rot (and What Survives)

Ruben Hassid spent six months teaching a Claude Cowork setup built on files and folders. He says 750,000 people trust his guides, that they pulled 20 million views, and that half of LinkedIn runs his system. On 1 July 2026 he told them all to delete it.

The retraction is worth reading in full: “I was wrong about Claude.” The opening line does not flinch:

“I spent months running Claude Cowork wrong. If you used my old guides, so did you.”

This is not a takedown. Retracting at that scale takes nerve, and the replacement guide is good. But the episode is the cleanest case study I’ve seen for a pattern every founder using AI will hit. Setups rot. The useful question is what survives.

What actually broke

Ruben’s account of the trigger: “I started giving workshops to enterprises. Teams with hundreds of people using Claude Cowork together. And my files + folder system broke.” His diagnosis:

“Files are impossible to maintain. You have a full-time job. Me too.”

Read that second quote twice. The folders did not fail on day one. They failed when nobody owned them.

Rot is the norm, not the exception

Setup advice rarely becomes wrong. It becomes irrelevant first, then counterproductive. Recent history is full of examples.

Chain-of-thought prompting was 2023’s flagship skill. By June 2025, Wharton’s Prompting Science Report 2 found it gave reasoning models marginal gains at 20 to 80 percent extra time cost. OpenAI’s own best-practices docs now say to avoid chain-of-thought prompts. Two years from gold rush to vendor deprecation notice.

Naive RAG, the chunk-embed-vector-database stack, was the default architecture of 2023 and 2024. Agentic search and just-in-time context loading displaced it. Retrieval still earns its keep on large, frequently updated corpora and citation-heavy work. What died was the copy-paste default.

Memory hacks went the same way. “Paste this about-me block into every chat” was standard advice until native memory shipped: ChatGPT in April 2025, Claude from October 2025.

And now the subject itself. Cowork launched in January 2026. Ruben’s folder screenshots are dated April 2026. Retracted 1 July 2026. Under six months, self-documented. His old infographic said “Pick Opus 4.7.” The new guide points at Fable.

The next walk-back is already scheduled

Skills plus Projects is the right setup today. But Ruben is already pointing at the next model, and Anthropic will reshuffle its primitives again. The deepest line in his piece:

“How Anthropic wants us to use their AI is different from how people actually use it.”

If vendor intent and user practice diverge, every third-party setup is a guess awaiting correction. Including his new one. Including mine.

That sounds bleak until you separate the layers. The setup layer is which primitive holds what, which model to pick, how to structure the files. Disposable. Plan to rebuild it every six months. The layer underneath is your voice, your workflows, your standards, your operating context. That layer compounds. Captured once, it ports across whatever ships next. When the primitives reshuffle, you re-pour the same material into new containers in an afternoon instead of starting from zero.

Setups are disposable. The brain is not.

One honest wrinkle

The new guide argues too much context kills creativity. The old one said to pile it on until failure was impossible. Both are right, for different jobs. Execution context compounds: feed it everything. Ideation context constrains: brainstorm with your full history loaded and you get variations of what you already do. The skill is curation, not quantity. An unowned folder gives you the worst of both.

Who this gap belongs to

The last page of Ruben’s post sells workshops for companies with 100+ people. He’s right about where the value sits: in the gap between generic advice and a system actually installed in a business. His retraction exists because his generic system met real teams and lost.

Founders at 5 to 50 people have the same gap, and nobody is pitching them workshops. They’re the ones copying setups from LinkedIn, running them with no owner, and finding out six months later that the whole thing needs rebuilding.

The move is not to find a better influencer setup. It’s to own the layer underneath: context captured from your actual business, installed where you and your team work, with a named owner and a maintenance loop. Then the next retraction costs you an afternoon, not six months.

That layer is what the AI Brain engagement builds for you.