Second Opinion

Input

A proposal, architecture, feature plan, refactor plan, implementation approach, review finding, or technical decision to sanity-check.
Use explicit input first; otherwise infer the proposal and target AI from context.
Safest default: detect available local AI CLIs and ask which one to use.

Clarify target and input. Identify the proposal to review and the AI CLI to ask. If either is missing, infer from context or ask one focused question.
Detect available CLIs. Check likely commands with command -v, such as claude, gemini, codex, cursor, opencode, and aider. Do not assume a CLI exists.
Verify invocation. Use the selected CLI's help output before running it. Prefer non-interactive prompt modes and avoid commands that open editors, mutate files, install packages, or start long-lived sessions.
Write the proposal. Create a temporary directory with mktemp -d and write the proposal, relevant constraints, and explicit questions to a file such as proposal.md.
Ask the other AI. Pass the proposal file content to the selected CLI. Ask for critique, missed risks, alternatives, and concrete changes to the proposal.
Capture feedback. Save the raw response in the same temporary directory. If the CLI fails, report the command, failure mode, and closest safe fallback.
Compare judgments. Separate useful objections from weak or irrelevant feedback. Do not accept the other AI's answer just because it came from another model.
Revise if warranted. Produce a final assessment, noting what changed, what did not, and why.