anti-gravity - SKILL.md Agent Skill

name: anti-gravity description: Use when the user asks to run Anti-Gravity CLI (agy) for Gemini-like code review, plan review, design critique, or large-context analysis. Anti-Gravity replaces the legacy Gemini CLI workflow.

Anti-Gravity Skill Guide

When to Use Anti-Gravity

When the user asks for Gemini-like analysis, review, planning, design critique, or broad context work.
When a second-model perspective is useful.
When the task can be expressed as a short, direct agy prompt.

Tested Behavior in This Harness

agy -p may produce an empty captured stdout inside the Pi/sandbox harness even when the model answered successfully. The reliable way to verify output here is:

run agy -p with a short exact prompt,
capture a log with --log-file,
read the created conversation id from the log,
inspect C:/Users/johno/.gemini/antigravity-cli/brain/<conversation-id>/.system_generated/logs/transcript.jsonl for the final model response.

The CLI stores conversation state under: C:/Users/johno/.gemini/antigravity-cli/

If the harness reports Access is denied while renaming files in conversations/, the response may still be present in the transcript. Prefer checking the transcript before retrying.

Core CLI Modes

agy -p "prompt" or agy --print "prompt" — run a single prompt and print/respond non-interactively.
agy --prompt "prompt" — alias for print mode.
agy -c or agy --continue — continue the most recent conversation.
agy --conversation <ID> — resume a previous conversation by ID.
agy -i "prompt" or agy --prompt-interactive "prompt" — start interactively and continue the session.
agy --sandbox — run with terminal restrictions enabled.

Workspace and Safety Flags

--add-dir <DIR> — add a directory to the workspace; repeat for multiple directories.
--log-file <PATH> — write CLI logs to a known path for debugging and transcript discovery.
--print-timeout <DURATION> — set print-mode wait time, e.g. 60s, 5m.
--dangerously-skip-permissions — auto-approve tool permission requests. Use only when explicitly appropriate.

Prompting Rules

Anti-Gravity follows instructions better when prompts are short and strict.

Use this style:

Do not use tools. Return exactly AGY_OK and nothing else.

For shell-like requests:

Return only the names of files and folders in the current directory, one per line, with no commentary.

Avoid vague prompts such as:

Can you look around and tell me what's here?

Reliable Pi/Harness Pattern

Use this when running from Pi/tooling and stdout is unreliable:

rm -f C:/Temp/agy-run.log
agy -p "Do not use tools. Return exactly AGY_OK and nothing else." \
  --print-timeout 60s \
  --log-file C:/Temp/agy-run.log

conv=$(grep -oE 'conversation=[0-9a-f-]+' C:/Temp/agy-run.log | tail -1 | cut -d= -f2)
read "C:/Users/johno/.gemini/antigravity-cli/brain/$conv/.system_generated/logs/transcript.jsonl"

Expected successful indicators:

process exits 0, and
transcript contains a MODEL entry with the requested answer.

Human Terminal Pattern

When the user runs Anti-Gravity directly in PowerShell, this can print normally:

agy -p "reply me with exactly the list of files in this directory and nothing more"

Common Use Cases

Direct exact response

agy -p "Do not use tools. Return exactly AGY_OK and nothing else."

Focused review

agy -p "Review this plan for missing risks. Return 5 bullets max."

Resume prior work

agy --conversation <conversation-id> -p "Continue the prior review. Focus only on deployment risks."

Multi-directory analysis

agy --add-dir src --add-dir docs -p "Map the feature flow across these directories. Return concise bullets."

Troubleshooting

Empty stdout in Pi does not necessarily mean failure; inspect the transcript.
Access is denied under conversations/ is a local storage/permission issue, not necessarily a model failure.
If the prompt requires tool use, expect Anti-Gravity to narrate tool intent unless instructed otherwise.
Use --log-file on every automated run so the conversation id and failure mode are recoverable.

Notes

Treat Anti-Gravity as the canonical successor to the old Gemini CLI skill.
Do not use deprecated Gemini CLI flags or model names.
Prefer exact, bounded prompts over broad exploratory ones.