gemini-consultant

stared 的头像作者 stared

使用 Google 搜索基础和视觉功能从 Gemini 3 Pro 获取第二意见。当您需要实时网络信息、验证事实、需要技术问题的不同视角、咨询其他 AI 模型或分析图像时使用。

📖 中文导读:使用 Google 搜索基础和视觉功能从 Gemini 3 Pro 获取第二意见。当您需要实时网络信息、验证事实、需要技术问题的不同视角、咨询其他 AI 模型或分析图像时使用。 以下正文为该技能的原始 SKILL.md 内容(英文),上方安装命令可直接使用。

name: gemini-consultant description: Get a second opinion from Gemini 3 Pro (gemini-3-pro-preview). Accepts text, code, and images as INPUT. Returns TEXT analysis, advice, and feedback. Use for code review, analyzing screenshots, UX feedback, debugging, architecture review, or web search. allowed-tools: Bash(uv:*), Read

Gemini Consultant

Get a second opinion from Google's Gemini 3 Pro (gemini-3-pro-preview) with real-time Google Search grounding.

CRITICAL: GEMINI IS BLIND. It knows nothing about your current session, file system, conversation history, or previous errors unless you explicitly force-feed it that data.

The "Dump Everything" Rule

The -c/--context argument is effectively unlimited. To get a useful answer, you must dump EVERYTHING relevant into the context.

  • Quantity: We are talking about 10, 100, or 500+ lines of context.
  • Scope: Include the file in question, all related files (imports, types, configs), the full error traceback, your current hypothesis, and a summary of what you have already tried.
  • History: If you are deep in a debugging session, dump the relevant parts of the conversation history into the context.

If you think a file is even tangentially related, INCLUDE IT.

Prerequisites

GEMINI_API_KEY environment variable must be set.

Usage

uv run /path/to/skills/gemini-consultant/consult.py "Detailed Question" -c "MASSIVE CONTEXT STRING"

Parameters

Parameter Description
question The prompt. Be specific. "Given error X and code Y (in context), why is logic Z failing?"
-c, --context THE DATA DUMP. Concatenate everything here. Code, logs, history, configs. Can be used multiple times.
-i, --image Image file paths (screenshots, diagrams).
--media-resolution low, medium (default), high, ultra_high.
--no-search Disable web search (pure code logic).
--thinking low or high (default). Use high for code.

Examples

BAD (Lazy = Failure)

# WRONG: Context is missing.
uv run consult.py "Why is this crashing?" -c "Error: 500 Internal Server Error"
# WRONG: Single file provided, ignoring dependencies.
uv run consult.py "Fix this function." -c "$(cat src/broken_file.py)"

GOOD (Exhaustive = Success)

NOTE: Real-world usages are usually 10-100x longer than the snippets below.

# CORRECT: Providing File, Imports, Types, and Errors
uv run consult.py "I am getting a RecursionError in the serialization logic. Analyze the relationship between the Model and the Schema." \
  -c "FILE: src/models.py
$(cat src/models.py)

FILE: src/schemas.py
$(cat src/schemas.py)

FILE: src/config/db_settings.py
$(cat src/config/db_settings.py)

ERROR LOG:
$(cat logs/error_dump.txt)

PREVIOUS ATTEMPTS:
I tried removing the circular reference in schema.py line 40 but it broke the API validation."
# CORRECT: Architecture review with full project scope
uv run consult.py "Review this authentication flow for security flaws." \
  -c "MIDDLEWARE: $(cat src/middleware/auth.ts)
MODELS: $(cat src/models/user.ts)
ROUTES: $(cat src/routes/auth.ts)
CONFIG: $(cat src/config/security.ts)
TYPES: $(cat src/types/express.d.ts)"

Web Search (No Code Context Needed)

uv run consult.py "What is the latest version of Next.js and what are its new features?"