IdeaWell

Project ideas from Hacker News discussions.

Claude Code reads AGENTS.md only when telemetry is on [fixed]

Original Article

Hacker News Discussion

📝 Discussion Summary (Click to expand)

1. Human error vs. AI‑generated code
Many commenters argue the issue is a straightforward human mistake—feature‑flag wiring tied to telemetry—not an AI‑generated bug.

“It seems to be exactly the opposite, a 'fully human error'… Nothing to do with AI patches at all.” – crazygringo
“Apologies again folks, this was a fully human error on my part…” – mpoteat
Others counter that any glitch is blamed on AI while successes are credited to it:
“LLMize the successes, humanize the failures.” – q3k

2. Feature‑flag/telemetry coupling as poor design
A recurring critique is that making a local‑file feature depend on a remote telemetry flag is indefensible and creates silent failures for users who opt out of telemetry.

“Reading a local file should never depend on a remote feature flag… This crosses a line.” – mgaldys4
“The whole impact seems to be that users who have telemetry turned off got this feature ~2 days later…” – HotHotLava
“If folks play around with it, I would love feedback… The mod is source available…” – mpoteat (acknowledging the flag‑based rollout)

3. Over‑engineering vs. simplicity
Debate centers on whether the new plugin/mod system (Mods, function hooks) is necessary for a trivial filename change or a justified step toward extensibility.

“The only difference between CLAUDE.md and AGENTS.md is the filename… Feels like this could have been a one line change.” – rickette / locknitpicker
“Having a kill switch for the change is a perfectly reasonable safety measure…” – epihelix
“This 'tiny feature' consists of hundreds of lines of code spread across 57 source files… It's the very definition of overengineering.” – bakugo

4. Trust, transparency, and user control
Users express concern about silent, remote‑controlled changes eroding trust, especially when there’s no opt‑out or warning.

“Even though Anthropic… could technically run anything on our system… breaks that trust for many users.” – dotancohen
“I do not like others to decide that they know what is best for me… and then force it on me without my consent.” – lukewarm707
“Apologies again folks… I should've found a better way to launch with a kill-switch.” – mpoteat (acknowledging the need for better user control)

🚀 Project Ideas

FlagLocal

Summary

  • A desktop utility that lets Claude Code users toggle feature flags via a local JSON config, eliminating the need for telemetry‑gated rollouts.
  • Core value proposition: restores user control over feature releases and prevents silent feature loss when telemetry is disabled.

Details

Key Value
Target Audience Claude Code power users who disable telemetry for privacy
Core Feature Reads/writes a local ~/.claude/flags.json that the tool patches into Claude Code’s feature‑flag checks at runtime
Tech Stack Tauri (Rust backend + frontend) or Electron with Node.js
Difficulty Medium
Monetization Hobby

Notes

  • HN users complained: “with telemetry off you don’t get those [feature flags]” (mpoteat) and “It’s a rollout artifact… we needed a way to turn this off remotely via feature flags if it broke something” – FlagLocal gives that control locally.
  • Enables discussion about user‑controlled feature flagging versus telemetry‑dependent rollouts in AI‑agent tools.

ClaudeConfigLoader

Summary

  • A small wrapper that intercepts Claude Code’s file‑read calls and decides whether to load CLAUDE.md, AGENTS.md, both, or a user‑defined priority order.
  • Core value proposition: removes confusion over which instruction file is used and lets users enforce their preferred behavior without overengineered plugin systems.

Details

Key Value
Target Audience Developers who maintain both CLAUDE.md and AGENTS.md or want deterministic loading
Core Feature LD_PRELOAD‑style interceptor (Linux/macOS) or Windows DLL hook that substitutes file paths based on a configurable rule file
Tech Stack C/C++ with libffi or Rust using windows-sys / libc; config via TOML
Difficulty Medium
Monetization Hobby

Notes

  • Commenters noted: “The only difference between CLAUDE.md and AGENTS.md is the filename… could have been a one line change” (locknitpicker) and “Having a kill switch for the change is a perfectly reasonable safety measure” (epihelix) – the loader provides that switch.
  • Sparks debate on simplicity vs. extensibility in agent harnesses and could be extended to support custom instructions per model.

OpenClaudeHarness

Summary

  • An open‑source, MIT‑licensed reimplementation of the Claude Code harness that exposes the full source, simplifies the plugin/mod system, and accepts community contributions.
  • Core value proposition: gives transparency, enables auditing, and reduces overengineering by providing a clean, minimal extension point.

Details

Key Value
Target Audience Enterprises and privacy‑conscious developers who want to inspect or modify the agent harness
Core Feature Drop‑in replacement for the proprietary Claude Code binary, with a clear plugin API (function hooks) and optional telemetry opt‑out
Tech Stack Go or Rust for the core; TypeScript for plugin interface; uses existing Claude API bindings
Difficulty High
Monetization Revenue-ready: Enterprise support tier ($X/month) + optional hosted version

Notes

  • HN users said: “It’s not open source… source available, since it’s published, but you legally can’t do anything with it” (wccrawford) and “Why not just make Claude code open source?” (glub) – OpenClaudeHarness directly addresses that demand.
  • Provides a practical platform for discussing open‑source AI agent tooling and could attract contributions from the community.

SkillSync

Summary

  • A daemon that maintains a centralized ~/.agents/skills directory and creates synchronized symlinks (or mirrors) for each agent’s expected skill folder (e.g., Claude Code’s .skills/synced, Cursor’s .skills).
  • Core value proposition: eliminates duplicate skill copies and keeps all agents’ skill sets consistent without manual linking.

Details

Key Value
Target Audience Power users who run multiple AI coding agents (Claude Code, Cursor, Codex, etc.) and maintain shared skill libraries
Core Feature Watches the central skills repo, updates agent‑specific folders via symlinks or file‑copy on change, with conflict‑resolution rules
Tech Stack Rust with notify crate for FS events; cross‑platform daemon
Difficulty Low
Monetization Hobby

Notes

  • User vorticalbox described the pain: “I made a ~/.agents/skills folder then ln both cursor and claude skills to point to that folder… claude uses .skills/synced… ends up duplicating skills” – SkillSync automates this.
  • Encourages discussion on interoperability standards for agent tooling and could evolve into a shared plugin marketplace.

Read Later