<< All versions

Skill v1.0.0

currentAutomated scan100/100
azmym/gemini-plugin/gemini-consult
──Details
PublishedSeptember 28, 2026 at 09:54 AM
Content Hashsha256:d3ef7e27f2acdd55...
Git SHAd55db48008c7
──Files
Files (1 file, 4.8 KB)
SKILL.md4.8 KBactive
SKILL.md · 105 lines · 4.8 KB

version: "1.0.0" name: gemini-consult description: Dispatch rule for consulting the five Gemini agents. Read this when deciding whether to get a Gemini second opinion and which agent to route to (researcher, validator, challenger, summarizer, reviewer). Enforces a one-consult-per-turn cap on manual dispatches.


Gemini Consult (dispatch rule)

This skill tells you (the main Claude) WHEN to consult Gemini, WHICH of the five agents to dispatch, and the cap that keeps consults cheap. It complements gemini-when-to-use (which routes among the seven capability skills for raw MCP calls); this skill routes among the five agents for structured, second-opinion work.

Gemini complements your reasoning. It does not replace it.

When to consult

Dispatch a Gemini agent when ANY of these is true:

  • Uncertain about a plausibly post-cutoff fact (library version, API shape,

CVE, pricing, changelog). Route to gemini-researcher.

  • A specific claim needs external verification before you rely on it. Route

to gemini-validator.

  • Two or more valid approaches are on the table and you want an

outside-the-context-window opinion. Route to gemini-challenger.

  • Finalizing a diff or PR for cross-cutting concerns (security, threading,

version drift, doc accuracy, dead code, complexity). Route to gemini-reviewer.

  • About to compact or hand off a long session and want the state preserved.

Route to gemini-summarizer.

Routing table

NeedAgent
Live facts / citationsgemini-researcher
Verify a specific claimgemini-validator
Challenge a decision / alternativesgemini-challenger
Generalist diff / PR reviewgemini-reviewer
Compress session / risk mapgemini-summarizer

Agents do not call each other (subagents cannot spawn subagents). You are the router: pick the one agent that matches and dispatch it.

Deferred Gemini tools in heavy-MCP sessions

When many MCP servers are connected, Claude Code defers MCP tools: the Gemini tool names appear in a system reminder, but their schemas are not loaded and a direct call fails. An agent must call ToolSearch (keyword query such as gemini search grounded) to materialize the schema before invoking the tool. The five agents already document this step in their own "Tool availability" section, so dispatch them normally. But if an agent's JSON comes back with confidence: "unavailable" / verdict: "unknown" and an error mentioning a missing Gemini tool, do NOT assume the server is down: first run /gemini-plugin:gemini-doctor. If doctor's check 2 passes but check 3 fails, the session is stale (restart Claude Code); if both pass, retry the dispatch.

The automatic design-review pass

Separately from the manual consults this rule governs, the plugin runs an AUTOMATIC, advisory design-review pass via hooks: whenever a design/plan artifact is written (a *-design.md spec, a *-plan.md, or a file under a specs//plans/ dir) or native plan mode exits, the plugin asks you to dispatch gemini-validator and gemini-challenger over it. That pass is part of the uncounted hook channel: it does NOT count against the one-consult-per-turn cap, and it is advisory (findings surface but never block). It is silenced by the same CLAUDE_PLUGIN_GEMINI_DISABLE_HOOKS=1 / brainstorm.off kill switch as every other hook. The blocking validator at native plan-mode exit is unchanged; only the added challenger and the file-artifact pass are advisory.

The one-consult-per-turn cap

At most ONE manual, rule-driven Gemini consult per turn, across all five agents. Each consult adds latency and tokens; the value comes from being selective.

This cap counts only the consults YOU choose to dispatch from this rule. The always-on hooks (session-start risk map, per-prompt grounding, plan validation, destructive-op challenge, pre-compact summary, done-claim validation) are a SEPARATE channel and are NOT counted against this cap. To quiet the automatic channel, the user can run /gemini-plugin:gemini-brainstorm-off.

Do NOT consult when

  • The user said "no Gemini", "skip validation", or "just do it".
  • The fact is project-internal (repo state, file contents, build output) and you

can Read it directly.

  • You already have direct evidence (a file Read, command output) that settles

the question.

  • The same question was already consulted earlier this turn.

Disagreement protocol

If a Gemini agent contradicts your draft:

  1. Do NOT silently defer.
  2. Surface BOTH positions to the user with their citations.
  3. The user picks. If the user has no preference, default to Gemini's position

when its citation is concrete and recent (post-2024); default to yours when Gemini's citation is stale or low-quality.

Do NOT use this skill for

  • Making raw Gemini MCP calls (use the capability skills via gemini-when-to-use).
  • Non-Gemini routing decisions.
All versions