An Agent Skill that offloads token-heavy, low-reasoning work to a cheap LLM via the Antigravity CLI (agy), so the bulk reading happens outside your main context and only a compact digest comes back.
Works with any agent that supports skills or custom instructions — Claude Code, Cursor, Codex, OpenCode, and others.
A good fit when all three are true:
| Condition | Example |
|---|---|
| (a) Large expected reading volume | Scanning all files in a repo, digesting a 50k-line log |
| (b) No deep reasoning needed per step | Listing endpoints, extracting env vars, checking for TODOs |
| (c) Only the final summary or list matters | You don't need raw file contents, just the aggregated result |
Typical cases
- Full-codebase inventories (API endpoints, env vars, TODOs, test coverage)
- Digesting large log or data files
- Whole-repo architectural summaries
- Mechanical compliance sweeps across many files
- Bulk web research — surveying many pages and returning a consolidated digest
For "where is X?" lookups, use the agent's native search or exploration tools instead.
Install and configure the Antigravity CLI:
npm install -g @antigravity/cli # or your platform's install method
agy auth login- The agent picks the best low-cost Gemini model (
agy models). - The agent builds a self-contained prompt that includes target paths, the question, output format, and an instruction to be exhaustive.
agyruns non-interactively, reads files or searches the web, and prints a digest.- Only the digest enters the agent's context — the raw file content never does.
Copy SKILL.md into your project's agent skill directory, or reference this repository from your agent's skill configuration.
| Agent | Skill directory |
|---|---|
| Claude Code | ~/.claude/skills/ |
| OpenCode | ~/.config/opencode/skills |
| Codex | ~/.codex/skills/ |
Once installed, the agent will invoke it automatically when conditions (a)–(c) are met. You can also trigger it explicitly:
"Use agy to inventory all API endpoints in this repo."
- Raise
--print-timeoutfor large jobs (default 5m, e.g.--print-timeout 15m). - Results come from a low-cost model: reliable for enumeration, weaker on judgment. Spot-check before building further work on top.
MIT