claude-to-agy: Instructions file for Claude Code

Claude-to-Antigravity Delegation Rules

You are equipped with a custom MCP tool called delegate_to_agy. You MUST use this tool to offload heavy, token-intensive tasks to save your own context window.

The routing gate - consume vs. operate

Before any read, search, or analysis, pick the mode:

• Operate-on - code you're about to edit. Read the file and edit it yourself - even if it's over 200 lines; don't delegate it. Never edit from a summary.
• Consume-and-discard - you want a verdict, not the bytes (audits, searches, reviews, git history, external research). Delegate to agy.
• Unsure which mode? If you don't yet know whether you'll edit, treat it as consume and delegate. Only keep it local once you've committed to editing it.
• Discovery is always consume. "Find all X across module Y", "which files use pattern Z", "where are the assert() calls" - these are grep/multi-file analysis even if editing immediately follows. Delegate to agy to get the file list, then read only the specific files you'll edit. The fact that editing comes next does NOT make the discovery step operate-on.

If a review leads straight into editing the same file, read it locally once - don't delegate then re-read it.

Terminal Command Delegation

BEFORE running ANY of these commands in a terminal, you MUST use delegate_to_agy instead:

• grep - always delegate. Flag combinations (-r, -A, -B, -C) make output size unpredictable regardless of -m.
• git diff / git log - delegate by default. Exception: when you can bound it small and need it inline - git log -n 5, git diff --stat, a git diff of the single file you just touched - run it directly.

NEVER run grep directly. NEVER run unbounded git diff / git log directly. This applies during ALL phases: planning, exploration, implementation, review.

Delegation Thresholds

Consume-and-discard work goes to agy per the gate above. On top of that, delegate when ANY of these conditions are met: