name: llm-routing
description: Reference for choosing between LLM tools (ask-llms, llm-council, remote-llm). Consult before querying multiple models.
user_invocable: false

LLM Routing

Decision guide for when to use which LLM querying tool.

Tool Selection

Decision Tree

Is this about proprietary/work code?
  └── YES → /remote-llm (craft prompt for local LLM)
  └── NO ↓

Is this an important decision with trade-offs?
  └── YES → /llm-council (5 models deliberate)
  └── NO ↓

Do you need multiple perspectives quickly?
  └── YES → /ask-llms (parallel queries)
  └── NO → Just use Claude directly

/ask-llms

Purpose: Quick parallel queries to multiple models via OpenRouter.

Best for:
- Comparing model outputs on same prompt
- Getting diverse perspectives without deliberation
- Draft review from multiple viewpoints
- Quick "second opinion" checks

Flags:
- --cheap — Use cheaper models (for low-stakes queries)
- No flag — Use frontier models (for important messages)

Example:

/ask-llms "Should I follow up with this recruiter now or wait?"

/llm-council

Purpose: 5 frontier models deliberate sequentially, each seeing previous responses, then judge synthesizes.

Best for:
- Career decisions (job offers, timing, strategy)
- Important outreach messages
- When you want consensus, not just comparison
- Complex trade-offs with no clear answer

Models: Opus 4.5, GPT-5.2, Gemini 3 Pro, Grok 4, Kimi K2.5

Example:

/llm-council "I have an offer at $93K from Capco. Should I accept or negotiate?"

/remote-llm

Purpose: Craft prompts for Terry to run on local/work LLMs when code can't be shared.

Best for:
- Proprietary bank code
- Work systems Terry can't paste
- Anything requiring Qwen3 at CITIC

Output: A well-structured prompt Terry can copy to the work LLM.

Environment Setup

# Required for ask-llms and llm-council
export OPENROUTER_API_KEY=...

# For some models in council
export GOOGLE_API_KEY=...
export MOONSHOT_API_KEY=...

Cost Awareness

Coding Tools (Claude Code vs OpenCode)

Separate from querying multiple LLMs — this is about which coding assistant to use.

When to Switch to OpenCode

Stay in Claude Code if context is already built — switch costs (re-explaining, re-reading) usually exceed savings.

Suggest OpenCode (GLM-4.7) when:
- New task without existing Claude Code context
- Weekly Claude Code quota running high (>70%)
- Bilingual projects (TC/SC/EN) — GLM's multilingual edge

GLM-4.7 Notes

• Terry has unlimited quota via Coding Max (valid to 2027-01-28)
• SWE-bench Multilingual: 66.7% — strong for bilingual
• Preserved Thinking: keeps reasoning across agentic turns

Quota Conservation

When Claude Code usage is high:
- Default to OpenCode for new tasks
- Shorter responses, fewer exploratory reads
- Skip optional verification unless critical

Related Skills

• /ask-llms — Parallel queries implementation
• /llm-council — Deliberation implementation
• /remote-llm — Local LLM prompt crafting
• /opencode-delegate — Delegate tasks to OpenCode