# Second Opinion

> Ask another local AI CLI for feedback on a proposal, architecture, feature plan, refactor, review, or technical decision, then compare the feedback and produce a revised assessment. Use when the user wants another model's opinion or asks to check a plan with Gemini, Claude, Codex, Cursor, or another installed AI tool.

- Skill: `hunvreus/second-opinion` (Agent Skill, multi-file: 2 files)
- Install (CLI): `npx skillmds@latest add hunvreus/second-opinion`
- Raw SKILL.md: https://api.skillmd.com/api/skills/hunvreus/second-opinion/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: hunvreus (https://skillmd.com/u/hunvreus)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/hunvreus/second-opinion

---


# 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.

## Workflow

1. **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.
2. **Detect available CLIs**. Check likely commands with `command -v`, such as `claude`, `gemini`, `codex`, `cursor`, `opencode`, and `aider`. Do not assume a CLI exists.
3. **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.
4. **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`.
5. **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.
6. **Capture feedback**. Save the raw response in the same temporary directory. If the CLI fails, report the command, failure mode, and closest safe fallback.
7. **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.
8. **Revise if warranted**. Produce a final assessment, noting what changed, what did not, and why.

## Output

- Target AI and CLI used
- Proposal file path and feedback file path
- Other AI's strongest points
- Points rejected or deprioritized
- Final assessment
- Revised proposal when useful

## Examples

- `second-opinion with gemini on this architecture plan`
- `ask Claude for a second opinion on the refactor`
- `get another model to critique this feature proposal`
- `compare our deploy plan with Codex`

## Guardrails

- Do not send secrets, credentials, private keys, tokens, customer data, or unnecessary proprietary context to another CLI.
- Do not run a paid, networked, or external AI CLI unless the user requested that tool or approved it.
- Do not let the other AI mutate the repository; use it for feedback only.
- Keep the prompt focused. Send the proposal and constraints, not the entire repo, unless explicitly requested.
- Treat second opinions as advisory evidence, not authority.