LLM=True

Dev tries to leash noisy build logs; commenters unleash chaos

TLDR: A developer trimmed noisy build logs to save AI token space, and users split between applauding, proposing a separate “summary bot,” and accusing LLMs of wasting tokens on their own. It matters because cutting clutter can make AI assistants cheaper, faster, and less chaotic during real dev work.

A human-coded rant just made AI fans bark: the author’s “LLM=True” experiment aims to save precious AI brain-space (called tokens) by muzzling noisy build logs in Turborepo when working with Claude. They cut the chatter with error-only output and an update-silencing switch, while Claude cheekily added “| tail -5” like a good dog dodging the mess. But when builds fail, tail grows… 10, 30… drama! The crowd brought popcorn. Supporters like bigblind cheered the “wait, logs cost tokens?!” epiphany, and Bishonen88 dreamed of popular tools shipping “quiet modes” by default. Then came the split: vorticalbox pitched using a separate helper bot to run builds and deliver a tidy summary, sparing the main chat from log spam. Meanwhile, keybored rolled in with a hot “stop counting pennies” take, arguing the AI should manage itself. The spiciest mood swing? canto’s “this is just the start”—a rant that LLMs themselves bloat conversations with overcautious steps, extra scripts, and no caching, burning tokens faster than the logs ever could. The vibe: part clever housekeeping, part “fix the dog,” part “fix the whole kennel.”

Key Points

  • Excessive build logs from Turbo in a TypeScript monorepo consume LLM context windows, with a single build producing ~750 tokens.
  • Turbo can be configured to reduce output by setting "outputLogs": "errors-only" in turbo.json.
  • Setting the environment variable "TURBO_NO_UPDATE_NOTIFIER=1" disables Turbo’s update notifier.
  • Claude Code can pipe build output through "tail" (e.g., tail -5) to limit lines shown and reduce context pollution.
  • On build failures, the agent increases tail length (e.g., -10 or -30) to capture enough error context without flooding the window.

Hottest takes

"I never considered the volume of output tokens, but yeah, I like this idea a lot" — bigblind
"Why are you counting pennies instead of getting the LLM to do it?" — keybored
"Absolutely no cache (except the memory which is meh) nor indexing whatsoever." — canto
Made with <3 by @siedrix and @shesho from CDMX. Powered by Forge&Hive.