BITPlan

⚠️ LLM-generated content notice: Parts of this page may have been created or edited with the assistance of a large language model (LLM). The prompts that have been used might be on the page itself, the discussion page or in straight forward cases the prompt was just "Write a mediawiki page on X" with X being the page name. While the content has been reviewed it might still not be accurate or error-free. ✍️🤖 Agent/Seth: Did you know that Anthropic does not listen?

This page is part of the Does not listen series, after Microsoft does not listen. Pain is measured on the IT Pain Scale from 0 (no pain) to 10 (showstopper).

Your rules are only suggestions

Picture customers of a billion dollar AI company setting a reminder: every two weeks, post a comment saying nothing — so that a bot does not delete their security bug report.

That is not a joke. That is https://github.com/anthropics/claude-code/issues/53223

You write your rules into AGENTS.md or CLAUDE.md. Do not touch the secrets. Ask before you act. Read the boot page first. The documentation recommends exactly this file for security policies and compliance requirements.

The agent reads your rules — and then does what it likes.

The rule scientists use

In plain words: if you write your rules into the rule file, the agent follows them.

In scientific hieroglyphs: 

∀a ∀r ( Agent(a) ∧ r ∈ rules(CLAUDE.md) → complies(a, r) )

We do not believe rules. We count. Our agent incident records, open as of 2026-10-18: 

982 records, 1,713 incidents
328 evidence ignored
318 rule substituted
180 instructions ignored

One record falsifies the rule. We have 982.

Pain: 8 utterly horrible — we had to build and maintain our own incident measurement system to work around a product that ignores what it is told.

A 2010 Mac Pro meets every requirement — and crashes

Anthropic's system requirements: macOS 13.0 or newer, 4 GB RAM, an x64 or ARM64 processor.

Our Mac Pro 5,1 from 2010: macOS 14.7.8, 32 GB RAM, Intel Xeon X5690 — x64. Every box ticked. Then Claude Code updated itself. Illegal instruction. Dead.

The rule scientists use

In plain words: if your Mac has macOS 13 or newer, 4 GB RAM and an x64 processor, Claude Code runs.

In scientific hieroglyphs: 

∀m ( Mac(m) ∧ macOS(m) ≥ 13.0 ∧ RAM(m) ≥ 4 GB ∧ CPU(m) ∈ {x64, ARM64} → runs(m, claudeCode) )

What nobody tells you sits three dependency levels down: 

Mac Pro ─hasCPU→ Xeon X5690 ─microarch→ Westmere ─lacks→ AVX2
claudeCode ─shippedAs→ bun binary ─embeds→ JavaScriptCore ─prebuiltFor→ Haswell ─requires→ AVX2

One machine falsifies the rule. And the details change with every release.

So we did it ourselves: built bun and WebKit from source, added a pending upstream JIT fix https://github.com/oven-sh/WebKit/pull/292 by another user, and published the whole recipe https://gist.github.com/WolfgangFahl/f7c6574a8fcf1d203b507da4340b4b23 Today Claude Code runs again on our 2010 Mac Pro.

Pain: 10 showstopper out of the box — still severe with the workaround, unless you enjoy compiling compilers.

How the roles handle your needs

All of them ignore you. In the end you are left alone with the problem — and that is frustrating.

Will my software give me bigger frustrations than I am willing to cope with? These days nobody can answer that top down any more. One simple predictor is left: does the company listen?

BITPlan listens

At BITPlan we spend time with listening. Because we can: we are not a billion dollar company with lots of hierarchy levels. You can talk to the owner just like that.