ask-the-peanut-gallery

name	ask-the-peanut-gallery
description	Asks multiple AI models the same question via Cursor agents and synthesizes their answers. Use when you want diverse perspectives on a codebase question, architecture exploration, or code review.

Ask the Peanut Gallery

Delegate a question to multiple AI models (GPT, Claude/Sonnet, Gemini Pro, Gemini Flash) running as Cursor agents, then synthesize their responses.

The cursor-agent-multi.py, cursor-agent-task.sh, and cli.sample.json files are in the same directory as this skill file. Find them by looking in the skill directory.

Prerequisites

The target workspace must have a .cursor/cli.json file that controls what the Cursor agents are allowed to do. If it is missing, tell the user and show them the path to cli.sample.json (in this skill's directory) so they can copy and customize it:

mkdir -p <workspace>/.cursor
cp /path/to/cli.sample.json <workspace>/.cursor/cli.json

The sample allows read-only access, git commands, coreutils, and common test/build tools. The allow list acts as a whitelist — only listed commands are permitted. Users should add project-specific shell commands (e.g. Shell(ninja -C build **), Shell(ctest --test-dir build **)) as needed.

Steps

1. Locate the scripts. Find the directory containing this skill's files and use the cursor-agent-multi.py script there.

2. Run cursor-agent-multi.py with the user's question. Default the workspace to the current working directory.

   /path/to/cursor-agent-multi.py \
     --workspace <WORKSPACE> \
     --task <short-kebab-case-name> \
     --agents '<AGENTS_JSON>' \
     "<THE QUESTION>"
   

   Pick a short descriptive --task name based on the question (e.g. vmvx-architecture, flag-review).

   The --agents flag takes a JSON array of agent configs. Each object must have name and model. Example with default models:

   [
     {"name": "gpt",          "model": "gpt-5.3-codex-fast"},
     {"name": "claude",       "model": "sonnet-4.6"},
     {"name": "gemini-pro",   "model": "gemini-3.1-pro"},
     {"name": "gemini-flash", "model": "gemini-3-flash"}
   ]
   

   Additional options (pass through from the user if specified):

   • --include-dir DIR — directory of .md files for {{PLACEHOLDER}} resolution
   • --timeout SECS — per-agent timeout (default: 480)

   Run cursor-agent --list-models to see all available models.

3. Read all output files. After cursor-multi finishes, it prints the paths. Read every output.md file.

4. Return a synthesis in the following format:

   Peanut Gallery:

   Synthesis

   Summarize the consensus across models. Note any disagreements or unique findings that only one model surfaced. Be specific — cite file paths, function names, PR numbers, etc. when the models provide them.

   Individual Responses

   For each model that produced output, include a section with the model name and the verbatim content of its output.md.

Important

• Do NOT modify any files in the workspace.
• If some models fail, still return results from the ones that succeeded and note which failed.
• If cursor-agent-multi.py itself fails (e.g., missing cli.json), report the error.
• Always include this line in every prompt sent to agents: "You are running non-interactively. No human will see your questions or reply. Never ask for clarification. Make reasonable assumptions and state them. If a tool call fails, try alternative invocations before giving up. Provide a complete answer no matter what."