Apple accidentally left Claude.md files Apple Support app

Original Article ↗ Hacker News Discussion ↗

Apple accidentally shipped an internal `CLAUDE.md` configuration file for Anthropic’s Claude AI inside the Apple Support app, revealing that some teams use external LLMs as part of their development workflow. Commenters debate how deeply AI-powered “vibe coding” is embedded at Apple, contrasting the company’s traditional reputation for careful UX and engineering with growing reliance on code assistants across the industry. The incident also sparks broader concerns about declining software quality, AI-generated “slop” and bots on social platforms, and best practices for managing agent instruction files like `CLAUDE.md` in source control and build pipelines.

Scope of the mistake

  • Apple’s Support app accidentally shipped Claude.md files, which are instruction/config files for Anthropic’s Claude.
  • The files appear to describe project structure, coding rules, and “don’t do X” guidance, not secrets like tokens or proprietary algorithms.
  • Most see it as an embarrassing process slip (forgot to exclude from the app bundle), not a security incident.

Apple’s use of Claude and AI

  • Cited reporting claims Apple runs custom versions of Claude on its own servers for internal tools and product development.
  • Some argue this fits Apple’s pattern: rent external models while sitting out the public “AI arms race,” then buy/build when prices collapse.
  • Others are skeptical of the reporting details and spin, and note Apple’s difficult B2B posture may have pushed it toward larger partners like Google for Siri/Gemini.

Siri, voice assistants, and LLMs

  • Many think Siri has stagnated and lags behind Alexa, Gemini, and especially ChatGPT’s voice mode in understanding and usefulness.
  • Counterpoint: for driving and home control, people prefer “voice-driven command-line” behavior—predictable, deterministic commands over chatty LLMs.
  • Reports on Gemini/Alexa+ are mixed: some call them “objectively better,” others say basic tasks (timers, routines, smart home actions) regressed, became slower, or more unreliable.
  • Hallucinations and non-determinism are seen as blockers for billions of users and safety-critical flows.

AI-assisted coding and “vibe coding”

  • Multiple commenters say nearly every large tech company is pushing AI coding hard; Apple engineers almost certainly use LLM tools too.
  • Concerns: AI-generated “slop,” tech debt, and review fatigue when humans must debug large, low-quality diffs from LLMs.
  • Strong agreement that production changes still need human review and process; using AI isn’t the problem, lack of rigor is.

Should Claude.md live in source control?

  • Many argue “yes”: agent instruction files are part of the codebase, akin to build configs, style rules, and documentation; they must be shared and versioned.
  • Others initially treated them like local IDE cruft, but were persuaded that shared AI behavior and review automation justify checking them in.
  • Consensus: they belong in the repo, but must be excluded from final builds.

LLM spam and community quality

  • Several note that replies on X and increasingly HN show “LLM smell,” likely farming engagement/karma.
  • This is seen as degrading discussion quality and pushing people toward smaller, curated communities.