antigravity-mcp
CLAUDE.md0 starsChanged 2 days agoMITView on GitHub
This repository ships three Claude Code skills that use the Antigravity CLI (agy, Gemini). This file is re-read from disk at the start of every session and after every compaction, so these rules hold for the whole session. The skill texts themselves can be dropped or truncated by compaction.
In this repository3 other files it gives its agents.
- antigravity-advisorSkill
- antigravitySkill
- dual-advisorSkill
The file
We split this CLAUDE.md into the rules it gives an agent
7 rules. About 0 are common, 2 are distinctive, and 5 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈2
Distinctive · estimate
This author's own choices. Worth a look.
5
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- could be automatic: a setting could enforce it
- Very specific: line 10
antigravity-advisor: read-only second opinion from agy on a plan, an approach or finished work. - Distinctive: line 11
Use when the user asks for Antigravity/Gemini as advisor.
- Very specific: line 19
Never run agy -p from memory.
- Very specific: line 25
Never trust the agent's report: Check git status, git diff and run the tests yourself before saying a delegated task is done.
- Very specific: line 31
Quota: Check agy -p "/quota" before long or repeated use, as the skills describe.
- Distinctive: line 48
No scratchpad listed: keep it in context only and never use a cross-session file.
- Very specific: line 54
One
agyprocess per conversation at a time: Do not consult the advisor while a background worker is resuming the same conversation: wait …
≈ “Common” is an estimate: at least 1 in 5 of the 28,029 CLAUDE.md files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Antigravity skills: standing instructions
This repository ships three Claude Code skills that use the Antigravity CLI (agy, Gemini). This file
is re-read from disk at the start of every session and after every compaction, so these rules hold for
the whole session. The skill texts themselves can be dropped or truncated by compaction.
The skills and when to use them
antigravity: delegate a bounded coding task toagyas a background worker (own git worktree, you verify the result). Use when the user asks for Antigravity, Gemini or agy to do work.antigravity-advisor: read-only second opinion fromagyon a plan, an approach or finished work. Use when the user asks for Antigravity/Gemini as advisor.dual-advisor: for a hard problem, Claude's built-in advisor and the Antigravity advisor answer independently, then Claude's advisor merges them. Use only when the user asks for both. Needs the built-inadvisortool (/advisor <model>).
Rules that always apply
- Load the skill before using
agy. Invoke it with the Skill tool even if you remember the commands: after a compaction the full text may be gone or cut, and the Skill tool restores it. Never runagy -pfrom memory. (agy modelsandagy -p "/quota"are fine on their own.) - Model and reasoning level are chosen by the user each time a new
agyconversation starts (ask: each-use): ask unless the request names them. TheConfigurationblock at the top of.claude/skills/antigravity/SKILL.mdand.claude/skills/antigravity-advisor/SKILL.mdholds only the suggested defaults (and theaskswitch). Do not hard-code a model name anywhere else. Flags are stored per role (worker, advisor) inagy-session.jsonand reused on follow-ups. - Never trust the agent's report. Check
git status,git diffand run the tests yourself before saying a delegated task is done. Never use--add-dir. Never run two workers in one directory. - Privacy. An advisor briefing is sent to Google: no secrets, no personal data, no raw transcript.
- Hooks are session-wide.
.claude/settings.jsonregisters the guards in.claude/hooks/*.sh(they checkagy -pcalls). Do not copy them back into a skill's frontmatter: they would run twice. Hook scripts keep LF line endings (.gitattributes). - Quota. Check
agy -p "/quota"before long or repeated use, as the skills describe.
One agy conversation per directory (session continuity)
Within one Claude Code session, one agy conversation per directory is shared by every role (prime,
worker, advisor, dual-advisor), so Gemini keeps the folder context. Flow the user wants: at session
start the antigravity skill primes the conversation (reads the folder, read-only); later the same
conversation does the coding and acts as the advisor, each call with its own model and rules.
- Verified: a conversation is bound to the directory it started in (resumed from another directory it kept its memory but wrote into the ORIGINAL directory). So the unit is one conversation per absolute directory. A worktree is a different directory: a new conversation WITHOUT the primed context, so prime it again or work in the main checkout.
- Verified:
--model,--effortand--settingsare per call, also when resuming. The model really switches (the target model's own quota bucket drops), and deny rules apply (a write was blocked). So the same conversation runs the worker on one model, the advisor on a stronger one, read-only. - State file:
agy-session.jsonin this session's scratchpad directory (the system prompt lists it; private to the session, survives compaction). Shape:{"<abs dir>":{"id":"<conversation_id>","flags":{"worker":"<model-flags>","advisor":"<model-flags>"}}}. No scratchpad listed: keep it in context only and never use a cross-session file. Prime uses the worker's flags (read-only), so the user is asked at most twice per session. - Before any
agytask or advisor call (not/quota): read the file. Entry for the directory you startagyin: pass--conversation <id>with this call's own flags. No entry: start a new one. - After the result is read (background run: after its notification): write the returned
conversation_id, and the flags used under that role. Empty ID (failed to start): write nothing. - One
agyprocess per conversation at a time. Do not consult the advisor while a background worker is resuming the same conversation: wait for it to finish. - Start fresh (drop the entry's
id, keep itsflagsso the user is not asked again; delete the whole entry only when the directory itself is gone) only when the user asks for a new agy session, the worktree is removed, or the conversation has grown too long (lastusage.input_tokensabout 200,000 or more) or its quota is low. Writing the ID back after EVERY call is manual: do not skip it. - Anchoring caveat: Gemini's earlier history (briefs that held Claude's plans, its own work) can colour
its later "independent" opinion. State it in the
dual-advisorreport, do not engineer around it. - The briefing privacy rule covers the whole conversation. Input grows each turn (9.5k to 47k tokens over four trivial turns), so keep follow-ups short.
Automatic advisor (opt-in)
automatic-advisor: off
Change off to on to make this standing behaviour: for a non-trivial task, load antigravity-advisor
and consult it before substantive work and again before declaring done (and when stuck), at most
once per moment. Every consult sends a briefing to Google and spends quota, so it stays off until
the user turns it on. With off, consult only when the user asks.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

