| name | consult-second-opinion |
| description | Use when the user asks for a "second opinion", "third opinion", "what does GLM/Gemini think", or wants another model to weigh in on code, design, or review. Routes to the consult plugin's CLI which calls z.ai (GLM-5.2) or Gemini directly via API. |
| triggers | ["second opinion","third opinion","ask glm","ask gemini","ask z.ai","what does glm think","what does gemini think","another model","consult zai","consult gemini"] |
Consult — Second Opinion Skill
This skill activates when the user wants another model family to weigh in. The plugin ships a small CLI (scripts/consult.mjs) that hits z.ai's GLM-5.2 or Google Gemini directly. No extra CLI binaries needed — just an API key in env.
When to use
- "Get a second opinion from GLM on this design"
- "Ask Gemini whether this migration is safe"
- "What would z.ai do here?"
- "Run a third-party review of this PR"
How to invoke
For free-form questions:
node ${CLAUDE_PLUGIN_ROOT}/scripts/consult.mjs ask --provider zai "<the question, with all needed context inlined>"
For a code review of the current branch (auto-gathers git diff):
node ${CLAUDE_PLUGIN_ROOT}/scripts/consult.mjs review --provider zai
node ${CLAUDE_PLUGIN_ROOT}/scripts/consult.mjs review --provider gemini --focus "security"
Use --provider gemini when the user explicitly asks for Gemini, otherwise default to zai.
Important context guidelines
- The remote model has no access to the local filesystem. Inline every piece of relevant code or context into the prompt argument. Don't reference files it can't see.
- Keep prompts focused. Long context dumps are fine when justified, but trim to what's load-bearing.
- For reviews, the script handles diff extraction — don't paste a diff manually.
Provider keys
Verify with /consult:config. If missing, ask the user to set:
ZAI_API_KEY for z.ai (GLM)GEMINI_API_KEY for Google Gemini
Output handling
Show the model's response verbatim. Surface its verdict (if any) at the top of your reply so the user sees it without scrolling. If it's a code review and the verdict is BLOCKER or NEEDS WORK, list the cited issues clearly.
