LLM Call

LLM Call

External LLM access. Only activates on explicit triggers.

Triggers

No trigger → Claude handles alone.

Auto-save: All requests automatically save to session with step-based folders for history tracking.

Why This Pattern Exists

The value is independent perspective, not review.

If Claude shows its draft to external models, they anchor on it. Instead:

1. Claude forms complete answer first
2. External models answer the same question independently (they never see Claude's draft)
3. Claude compares all answers afterward

External models cannot: search web, use tools, see files, or access conversation history. Claude must include all relevant context in the query.

Workflow - Follow STRICTLY

Input: ===QUERY=== (required), ===DRAFT=== (optional), ===PROBE=== (probe only)

Draft: If Claude has answered the question, Claude SHOULD NOT include this section in council phase to save context window. Claude can pass the draft if the user invoked @crossref afterall.

Single Model (@gpt, @gemini, @grok, @qwen)

cli.py -m single -M gpt << 'EOF'
===QUERY===
Question + context
===DRAFT===  
Claude's answer
EOF

Council (@council)

cli.py -m council << 'EOF'
===QUERY===
Question + context
===DRAFT===
Claude's answer
EOF

Add -c for confidence ratings.

Probe (@probe)

Follow-up with auto-context from ALL previous steps.

cli.py -m probe << 'EOF'
===QUERY===
Explain more about [point]
===PROBE===
@gpt
EOF

Auto-gathers history → sends to model → saves to new step.

Crossref (@crossref)

Purpose: Each model sees what ALL others said and comments on their responses.

Crossref requires Claude's draft. Two ways to provide it:

1. If you already included ===DRAFT=== in the council step:

   cli.py -m crossref

2. If you only sent the query in council (no draft):

   cli.py -m crossref << 'EOF'
   ===DRAFT===
   Claude's answer here
   EOF

What each model receives:

• Original question
• Their own previous answer (if they had one)
• Claude's draft (if available)
• All OTHER models' responses

ALL 4 models are always invoked, even if one failed in the council step. A model that failed earlier can still comment on others' responses.

Session: Auto-saves to /tmp/sessions/s_TIMESTAMP/1/, /2/, etc. Each step folder contains .md files (query, draft, gpt, gemini, grok).

Script Reference

Flags:

• -M model (gpt/gemini/grok/qwen)
• -c confidence mode
• -S session ID or new (optional)

The -c Flag (Confidence)

What it does: Asks each model to rate its confidence and explain what would change its answer.

When to use it:

• Factual/analytical questions where certainty matters
• To surface what evidence each model is relying on

The -S Flag (Session)

Options:

• -S new — Force create a new session (useful when starting a new topic)
• -S <session_id> — Use a specific session (e.g., -S s_20250101_120000_1234)
• (omit) — Auto-use current session, or create if none exists

When to use -S new:

• Starting a completely new topic/question
• Want to keep previous session separate