limite

limite

Claude Code runs every task on the session model. On Opus that includes finding a file, reading a log or renaming an import, and each of those eats the usage limit. The built-in Explore sub-agent inherits the session model too, so code search also runs on Opus. This skill moves the easy work to Haiku (model: haiku, always the latest Haiku), which costs a fraction of Opus per token.

What the user typed after /limite: $ARGUMENTS

Scope

• Default: user scope, every project. Agents go to ~/.claude/agents/, the note to ~/.claude/CLAUDE.md.
• --project (or the user asks for this project only): agents go to .claude/agents/, the note to ./CLAUDE.md at the repo root.

Steps

1. Read templates/Explore.md and templates/chores.md from this skill's folder.
2. For each one, check the target agents folder (create it if missing):
   • Missing: write it.
   • Present and identical: skip it.
   • Present and different: show the diff and ask before replacing it. Never overwrite silently.
3. Read templates/CLAUDE-note.md. If the target CLAUDE.md already has a ## Models section (or the same rules in another language), leave it. Otherwise append the section, translated to the language the rest of that CLAUDE.md is written in. Create the file if it does not exist.
4. Tell the user what was written, with the paths, then:
   • Restart Claude Code (or open /agents) so the new agents load.
   • Check /agents: Explore and chores must both show Haiku.
   • The session itself stays on Opus. Only the delegated work moves to Haiku.

Notes

• A user agent named Explore replaces the built-in one, that is the point of the first template.
• Haiku is fine for search, reading and one-line edits. Anything that needs judgment stays on Opus: the chores agent stops itself when a task touches more than 2 files or needs a design decision.
• Anthropic does not publish how much each model weighs in a subscription limit, only that Opus uses much more of it than Sonnet, and Sonnet more than Haiku. Do not promise a number.