Gemini 3 prompting
How to write effective prompts when delegating to or reviewing with agy (Gemini via the Antigravity CLI). Use when constructing an agy-bridge prompt for --type search/code/analysis/review/implement, when a delegated result comes back off-target, or when writing an adversarial review brief for Gemini. Triggers on: delegate to agy, ask Gemini, agy prompt, second opinion, review with Gemini.From its SKILL.md
npx -y skills add davdittrich/delegate-agy --skill gemini-3-promptingAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
2 things to look at
- no licenseNo license file was found in the repository. Code published without one is not open source by default, so using it at work is a question for whoever answers licensing questions where you are.
- 20 stars20 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
SKILL.md
5.0 KB, as published. Nobody here has run it
Prompting Gemini 3 through agy
You drive agy (Antigravity, Gemini) in print mode through agy-bridge. Print mode is headless: one prompt in, one result out. The agent cannot stop to ask a clarifying question, so the prompt you send is the entire brief. Write it like a work order for a fast, literal junior engineer.
The prompt travels via stdin; agy reads files itself from its workspace (--type code/analysis/review/implement grant read access). So describe what to do, point at where to look, and let agy read.
How Gemini 3 behaves — and how to prompt for it
- It follows instructions literally. "Fix the bug" → it fixes a bug its own way. "Make
parseDatereturnnullon empty input and add a test" → exactly that. Spell out the target behavior, not the vibe. - It is terse by default. Gemini 3 gives direct answers and skips narration. If you want a written plan or an explanation, ask for it explicitly.
- It plans and decomposes well. Give it the goal and the constraints; let it own the how. Over-scripting each step fights the model.
- Long context, but order matters. Put the data/code/diff first, then your instruction last. Anchor the ask to the material ("Based on the diff above, …").
- Negative constraints go LAST. "Do NOT touch the public API", "do NOT reformat unrelated files" — Gemini 3 can drop a negative constraint that appears too early in a long prompt. Put the don'ts at the very end. (This is why
--digestappends its contract last.) - One markup style. Markdown headings or simple labels are enough. Don't mix XML tags and Markdown in the same prompt.
Model selection is a flag, not a sentence. agy exposes --model; the bridge auto-selects per --type (search→Flash, everything else→Pro). Never write "use Gemini Pro" inside the prompt — pass --model "<exact name>" instead (agy models for names).
A solid delegate prompt has five parts
- Goal — one sentence, the outcome you want.
- Acceptance criteria — how you will know it's done (tests pass, endpoint returns X, build green). Turns a vague request into a checkable one.
- Where to look — the files, dirs, or modules that matter. Saves agy a blind search.
- Scope boundaries — what NOT to touch (other modules, public API shape, unrelated formatting). Put these last.
- Output expectation — code change only, or also a short summary of what changed and why. Add
--digestwhen you only want the compressed findings.
Keep it tight. A focused 8-line brief beats a 40-line essay; over-stuffed context buries the actual ask.
Choosing the --type (and what it may touch)
| Use | --type | When |
|---|---|---|
| Web-grounded lookup with source URLs | search | "latest", pricing, release notes, changelogs |
| Read-only reasoning over code | code / analysis | "explain how X works", root-cause, large-file analysis — no writes |
| Adversarial second opinion on a diff | review | independent cross-model critique; read-only, adversarial framing |
| Actually write files | implement | agy reads + writes; still no shell exec |
If unsure whether a task should write, start read-only (code) to get the plan, then re-run implement to apply it. Always verify agy's output yourself — a different model has different blind spots, and it can be confidently wrong.
Examples (on this bridge)
# Investigation — data-first, question anchored, read-only
echo "Given src/auth/*.ts, explain how token refresh is triggered and where it can race." \
| agy-bridge --type analysis
# Adversarial review — the diff is the material, the ask is last, negative constraint LAST
git diff main...HEAD | { cat; echo; echo "Review the diff above for correctness and edge cases. \
Be skeptical. Do NOT comment on formatting."; } | agy-bridge --type review
# Delegation with the 5-part brief + digest (compressed reply)
agy-bridge --type code --digest -- "Goal: list every retry site. \
Acceptance: a table of file:line + backoff strategy. Where: src/net/**. \
Output: digest only. Do NOT include unrelated files."
Limits to be honest about
- Print mode won't ask questions. Ambiguity becomes a guess — front-load the detail.
- Empty reply = failure, not silence. agy exits 0 with empty stdout on quota
RESOURCE_EXHAUSTED (429); the bridge/shim now fail loud (exit 3) with the reason. Wait for reset — re-prompting won't help. - Auth is the user's job. OAuth via Google account, no API key.
See also: skills/agy-delegate/SKILL.md for when to reach for the bridge and its type/model routing.