跳至正文
OpenAgentSkill
浏览已开放 · 目录记录日期 2026-11-21 · 自动更新和账户服务暂时关闭。
搜索与研究

agy-prompting

Internal helper — how to tighten a user request into a sharp prompt for the Antigravity CLI (agy / Gemini 3.x with native web search and agentic tools)

MarcosNahuel · 已有静态检查记录

安装与使用

运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。

npx skills add MarcosNahuel/antigravity-plugin-cc --skill agy-prompting

技能说明


name: agy-prompting description: Internal helper — how to tighten a user request into a sharp prompt for the Antigravity CLI (agy / Gemini 3.x with native web search and agentic tools) user-invocable: false

Prompting agy (Gemini 3.x via Antigravity CLI)

Use this skill ONLY inside the antigravity:agy-rescue subagent when shaping the text you pass to agy --print. Do not use it to do independent reasoning, file inspection, or research yourself.

What agy is good at

  • Native web search with citation grounding (real URLs, not hallucinated).
  • Long-context reasoning over multiple sources.
  • Following structured-output instructions (markdown tables, JSON, frontmatter).
  • Tool calling for code generation and file edits when run with --dangerously-skip-permissions.

What agy is worse at than Claude

  • Following complex multi-turn workflow rules. Keep the prompt single-shot and self-contained.
  • Inferring implicit context — if a constraint matters, write it down.
  • Localized linguistic nuance — explicitly ask for the target dialect ("Spanish (Argentina)", "Brazilian Portuguese", etc.) when the output language matters.

Prompt shape that works

<TASK in one sentence>.

Rules:
- <constraint 1>
- <constraint 2>
- ...

Output format:
<exact markdown / json schema you want>

<DATA / TOPIC>:
<the actual content>

Keep the rules to <= 8 bullets. More than that and Gemini starts dropping the tail.

Things to ALWAYS include

  • Output language when relevant.
  • Whether citations are required and the citation format ([N] style works best).
  • The exact output structure with markdown headers.
  • An anti-hallucination clause ("if you could not find a solid source, say so explicitly" / "mark weak claims as [WEAK EVIDENCE]").

Things to AVOID

  • Multi-step "first do X, then do Y, then call me back" instructions. agy in --print mode is single-shot; chain steps from the caller side instead.
  • Demanding more than ~20 sources unless the user asked for it — quality drops fast past that.
  • Free-form "be creative" — for research, structure is what makes the output usable.
  • Mixing rescue work (code edits) and research (web search) in the same call. Pick one.

Routing flags reminder

These are CALLER-SIDE controls, never include them in the agy prompt text:

  • --print — non-interactive
  • --dangerously-skip-permissions — auto-approve tool use
  • --print-timeout 20m0s — high-intensity research needs long timeouts
  • --continue — resume the most recent agy conversation
  • --conversation <id> — resume a specific past conversation

--model is version-dependent (corrected v0.6.1): early agy 1.0.0/1.0.1 reject it with flags provided but not defined: -model, but agy 1.0.5+ accepts --model "<name>" (e.g. --model "Gemini 3.1 Pro (High)", community-confirmed on issue #76). Since the installed version isn't known at call time, this plugin defaults to NOT passing --model (cross-version-safe — agy uses its configured default). Pass it only when the caller explicitly asked for a model AND agy --version is ≥ 1.0.5. Either way, --model is a CALLER-SIDE flag — never put it in the prompt text.

When in doubt

Default to the templates baked into agy-rescue.md for research mode. Only deviate when the user explicitly asks for a different shape.