argue

name	argue
description	Adversarial Claude-vs-Gemini debate tool. Claude and Gemini form independent positions, then debate back-and-forth until convergence or the round cap. Claude participates as a genuine debater — not a summarizer. Produces a battle-tested synthesis written to ~/.claude/arguments/finals/<slug>.md. Use when the user says "/argue <topic>", "/argue plan <topic>", "/argue bug <symptom>", or "/argue tech <question>".
args	[{"name":"args","type":"string","description":"Topic and optional flags: --rounds N, --paths file1,file2, --seed find_bug|plan|audit, --model <id>, --history.\n"}]

Argue Skill Invoked

User has requested: /argue {{args}}

Steps

1. Parse arguments

Parse {{args}} into:

• topic_type: if args starts with plan , bug , or tech , extract it and strip from topic. Default: general.
• topic: all text before any -- flag (after stripping topic_type prefix). Required. If empty, call:

  AskUserQuestion: "What should Claude and Gemini debate?"
  

  Wait for the user's response before proceeding.
• max_rounds: value after --rounds (integer, 1–8). Default: 4.
• context_paths: comma-separated file paths after --paths. Split into a list.
• seed_tool: value after --seed (find_bug, plan, or audit). Default: null.
• model: value after --model. Default: null.
• history_flag: true if --history appears in args.

2. Load context

If --paths provided: Read each file using the Read tool. Truncate to 300 lines each. Collect into a context block. Track any paths that don't exist as skipped_paths.

If --seed provided: Load the appropriate Gemini tool via ToolSearch and call it:

• find_bug → mcp__gemini__find_bug
• plan → mcp__gemini__plan
• audit → mcp__gemini__audit

Store the seed output for inclusion in opening positions.

3. Opening positions (parallel)

Form two independent opening positions — neither model sees the other's take first.

Claude's opening: Reason about the topic yourself (considering any seed analysis and file context). Form a clear, specific position with supporting arguments. Write it down as your Round 1 position.

Gemini's opening: Call analyze with:

• messages: [{"role": "user", "content": "<topic + context + seed if any>. Form a clear position on this topic with specific supporting arguments."}]
• system_instruction: the debate system instruction (see §Gemini System Instruction below)
• model: the parsed --model value, or omit

Record both positions. Initialize conversation_history as a list for the history file.

4. Debate loop

Starting from round 2, up to max_rounds:

1. Claude counters Gemini: Read Gemini's last position. Form a genuine counterargument — find the weakest point, introduce a new angle, or concede specific points that are stronger than your position. This must be real reasoning, not keyword extraction.

2. Send to Gemini: Call analyze with the full conversation history so far (alternating user/model roles, where "user" = Claude's turns, "model" = Gemini's turns), plus Claude's new counterargument as the latest user message. Include the debate system instruction.

3. Gemini responds: Record the response.

4. Convergence check: As a participant, evaluate honestly:

   • Are you repeating arguments? Is Gemini?
   • Do you genuinely disagree with Gemini's last point, or are you reaching for objections?
   • If you agree with Gemini's position and have nothing substantive to counter → mark as converged.
   • Track consecutive convergence count. If converged for 2 consecutive evaluations → stop the loop.

5. Record: Add both turns to conversation history.

If max_rounds == 1: skip the loop entirely — just use the opening positions.

5. Synthesize

Write a final synthesis based on having actually participated in the debate:

• If converged:

  Agreed position: <1–3 sentence summary of the convergent position, incorporating the strongest reasoning from both sides>

• If not converged:

  Positions not fully reconciled after N round(s). Key tension: Strongest argument (Claude): <your best point that Gemini didn't fully counter> Strongest argument (Gemini): <Gemini's best point that you couldn't fully counter> Best available synthesis:

6. Save outputs

Slug: topic lowercased, spaces → hyphens, non-alphanumeric stripped, truncated to 40 chars.

Finals file (always written):

• Path: ~/.claude/arguments/finals/<slug>.md
• Content:

  # Argue: <topic>
  **Type**: <topic_type>  **Rounds**: <rounds_run>  **Converged**: <yes/no>
  **Date**: <YYYY-MM-DD>
  
  ## Synthesis
  
  <synthesis text>
  

• Overwrite if it already exists.

History file (only if --history flag was set):

• Path: ~/.claude/arguments/history/<slug>-<topic_type>-<YYYYMMDDTHHMMSS>.md
• Content: full transcript — each message formatted as **<MODEL NAME>**: <content>, separated by ---. Use "Claude" and "Gemini" as the model names (not "user"/"model").
• After writing, list all files in ~/.claude/arguments/history/ matching <slug>-*, sort by mtime descending, delete any beyond the 5 most recent.

Plan handoff (only if topic_type == "plan"):

• Write the synthesis to plans/<whimsical-adjective-gerund-noun>.md (check existing plans/ to avoid collision).
• Then ask:

  "Argument complete. Run /draft-plan to wire this into a story, or keep as standalone?"

• If the user says yes: invoke /draft-plan pointing at the plan file.

7. Report

Print the synthesis inline. Note:

• Path written to finals
• Path written to history (if --history)
• Any skipped_paths

---

Gemini System Instruction

Use this as the system_instruction parameter for every analyze call in this debate:

You are in an adversarial technical debate with another AI model. Your job:
- Form and defend genuine positions based on evidence and reasoning
- Concede specific points when the counterargument is stronger — don't defend weak positions
- Introduce new angles, not restate old ones
- Be specific: name trade-offs, cite patterns, reference concrete scenarios
- If you agree with the other side's point, say so explicitly and explain why it changed your view