---
name: gemini-prompting
description: Internal guidance for composing Antigravity (Gemini) prompts for coding, review, diagnosis, and research tasks inside the Antigravity Claude Code plugin
user-invocable: false
---

# Prompting Gemini through Antigravity

Antigravity wraps Google's Gemini model in a CLI. The model is verbose by default and tends to over-explain. To get great rescue/task output, follow these rules.

## 1. Be explicit and structured

Gemini works better when the request is well-structured. Use a header + numbered steps when possible:

```
TASK
Investigate why <X> fails on macOS but works on Linux.

CONTEXT
- Reproduces with `npm test` in /Users/.../my-project
- Failing test: src/utils.test.ts → "should hash file"
- Suspected root cause: BSD `sha256sum` not installed on macOS

WHAT I WANT
1. Root-cause confirmation (read the code, run the failing test if possible)
2. Concrete fix (with file diff)
3. Verification command
```

## 2. Tell it what NOT to do

Gemini tends to add tons of context, options, and alternatives. To keep responses tight:

```
CONSTRAINTS
- Do not propose multiple solutions — pick the best one.
- Do not explain what the code does at length. Focus on the fix.
- Do not write a README/CHANGELOG/migration guide.
- Output max 200 words of prose + the diff.
```

## 3. Provide file paths explicitly

Gemini has its own scratch workspace at `~/.gemini/antigravity-cli/scratch/`. To make it work on the user's actual repo, **always provide absolute paths** in the prompt:

```
Work in /Users/olivierdupont/Desktop/diversclaude/my-project/
Read src/utils.ts and src/utils.test.ts.
Apply the fix to src/utils.ts.
```

If the path is outside the default workspace, add the routing flag `--add-dir /path/to/project`.

## 4. Diagnosis vs. fix

Distinguish clearly between:
- **Diagnosis**: "Investigate and report. No edits."
- **Fix**: "Apply the fix directly to the files."
- **Plan**: "Write a plan but don't execute it."

Default behavior of `--dangerously-skip-permissions` is to allow file edits. If you want read-only behavior, say so explicitly in the prompt.

## 5. Resume vs. fresh

Gemini's `--continue` flag resumes the most recent session. Use it (via the runtime's `--resume` flag) when:
- Following up on a previous rescue
- Iterating on a draft fix
- Asking for more depth on the same investigation

Use a fresh session when:
- Completely new task, unrelated to previous
- The previous context is stale or wrong-headed

## 6. Antipatterns

- **Vague**: "Fix this bug" — what bug? where?
- **Too many sub-tasks**: keep one focused task per `task` call.
- **Asking for opinions**: Gemini's opinions are bland — ask for actions.
- **Trusting blindly**: ALWAYS read the resulting diff before committing.

## 7. Recommended boilerplate

```
TASK: <one-sentence outcome>

CONTEXT:
<3-8 bullets of what's known, file paths, what's been tried>

DELIVERABLE:
<file diffs, verification command, or specific answer>

CONSTRAINTS:
- Output max <N> lines.
- Do not modify files outside <list>.
- Do not refactor anything outside the immediate fix.
```