Debate

Context modes — truth table

Rationalizations (why you must NOT skip steps)

Phase 0 — Setup

1. Parse the subject and detect or use the specified --preset
2. Determine context mode: hybrid (default) or all-fresh (if --all-fresh flag)
3. Define 3 angles with strong, genuine opposition
4. Extract the project pitch: read PILOT.md and use the first non-heading, non-empty line as the one-liner. If PILOT.md doesn't exist or has no usable line, use the project name from pyproject.toml or CLAUDE.md title.
5. Assign the lone wolf: use the preset's default wolf persona (see Presets section). You may reassign if the subject demands a different outsider perspective.
6. Announce the setup to the user:

Debate: {subject}
Preset: {preset}
Mode: {standard|deep}
Context: hybrid (2 insiders + 1 lone wolf) | all-fresh

Angles:
  A — {name}: {one-line perspective} [insider]
  B — {name}: {one-line perspective} [insider]
  C — {name}: {one-line perspective} [lone wolf]

Launching Phase 1...

Choosing angles

Good angles create IRREDUCIBLE TENSION — not just "for/against" but three perspectives that cannot all be satisfied simultaneously. Examples:

• Architecture: Puriste (correctness) / Pragmatique (ship fast) / End-user (simplicity)
• Strategy: Optimiste (opportunity) / Sceptique (risk) / Outsider (what would a competitor do?)
• Data: Statisticien (rigor) / Ingénieur (performance) / Business (ROI)

If using a preset, start from its default angles but adapt to the specific subject.

Phase 1 — Independent arguments (PARALLEL)

Launch 3 agents in parallel using the Agent tool. Each agent receives:

Insider prompt template (Agents A, B — or all agents if --all-fresh)

You are Agent {A|B} in a structured debate.

## Your angle: {angle_name}
{angle_description}

## Subject
{subject}

## Domain context
{preset_attack_taxonomy}

{IF hybrid mode (not --all-fresh):}
## Project context
{contents of CLAUDE.md rules section}
{contents of GUIDELINES.md relevant sections}
{contents of PILOT.md current state}
{END IF}

## Instructions

Build the strongest possible case FROM YOUR ANGLE. You are not trying to be
balanced — you are trying to be RIGHT from your perspective.

Structure your response as:

### Position
State your core thesis in 2-3 sentences.

### Arguments (ranked by strength)
For each argument:
- **Claim**: what you assert
- **Evidence/reasoning**: why it holds
- **Implication**: what follows if this is true

### Anticipated objections
What would the other perspectives say? Pre-empt their strongest counter.

### Red lines
What would make you CHANGE YOUR MIND? (If nothing — you're not arguing
in good faith. There must be at least one falsifiable condition.)

Keep it focused — 3-5 arguments maximum, depth over breadth.

Lone wolf prompt template (Agent C — hybrid mode only)

You are Agent C (the lone wolf) in a structured debate.

## Your angle: {angle_name}
{angle_description}

## Subject
{subject}

## Project pitch
{one-line project pitch extracted from PILOT.md}

## Domain context
{preset_attack_taxonomy}

## Instructions

You know almost NOTHING about this project — only the pitch above.
This is intentional. Your value is in seeing what insiders cannot.

Build the strongest possible case FROM YOUR ANGLE, reasoning from
first principles. Challenge assumptions that project insiders take for granted.

Structure your response as:

### Position
State your core thesis in 2-3 sentences.

### Arguments (ranked by strength)
For each argument:
- **Claim**: what you assert
- **Evidence/reasoning**: why it holds
- **Implication**: what follows if this is true

### Anticipated objections
What would the other perspectives say? Pre-empt their strongest counter.

### Red lines
What would make you CHANGE YOUR MIND? (If nothing — you're not arguing
in good faith. There must be at least one falsifiable condition.)

Keep it focused — 3-5 arguments maximum, depth over breadth.

IMPORTANT: Do NOT include any information about the other agents' perspectives in Phase 1 prompts. Context isolation is critical to avoid herding.

Phase 2 — Cross-critique (PARALLEL)

Collect Phase 1 outputs. Launch 3 agents in parallel. Each agent receives the ARGUMENTS (not reasoning) of the other two agents.

Agent prompt template (Phase 2)

You are Agent {A|B|C} in a structured debate, Phase 2 (cross-critique).

## Your angle: {angle_name} (same as Phase 1)

## Your Phase 1 position
{agent_own_phase1_output}

## Agent {X}'s arguments
{agent_X_phase1_arguments_only}

## Agent {Y}'s arguments
{agent_Y_phase1_arguments_only}

## Instructions

Review the other agents' arguments. For each:

0. **Steelman** — before critiquing, reformulate the argument in its STRONGEST
   form. Show you understand it at its best. No strawman attacks.

1. **Concede** — which arguments are strong even from your perspective?
   Be honest. Refusing to concede anything is a sign of bad faith.
2. **Contest** — which arguments do you dispute? Give specific counter-evidence.
   Do not just say "I disagree" — say WHY with reasoning.
3. **Expose** — what are they MISSING? What blind spot does their angle create?

Then:

4. **Update** — has your position shifted? If yes, how? If no, why not?
   State your updated thesis.

IMPORTANT: You MUST contest at least one argument from each other agent.
If you genuinely find nothing to contest, explain what would need to be
true for the argument to fail. Agreeing with everything is not cross-critique.

{IF agent is lone wolf AND insiders reference project specifics:}
Engage with insider arguments on their reasoning, not their project knowledge.
Do not dismiss them because "you don't know the codebase" — challenge the LOGIC.
{END IF}

{preset_cross_critique_guidance}

Phase 2b — Feedback loop (DEEP MODE ONLY)

After Phase 2, check for convergence:

Convergence test: Compare the 3 "Updated thesis" statements.

• If 2+ agents converge on the core recommendation → proceed to Phase 3
• If all 3 remain divergent → run ONE more round of cross-critique (Phase 2b) with the updated positions. Then proceed to Phase 3 regardless.

Maximum: 1 feedback loop. Do not iterate further — diminishing returns.

Phase 3 — Synthesis (ORCHESTRATOR)

You (the orchestrator) now have all Phase 1 and Phase 2 (and 2b) outputs. You also have the FULL project context.

Step 1: Relevance check (lone wolf arguments)

The lone wolf debated with minimal project context. Some arguments may be:

• Irrelevant: valid in general but not applicable given project constraints
• Already decided: the project has an existing ADR or convention that settles this
• Off-scope: the debate drifted to adjacent topics

For each lone wolf argument, assess:

• Relevant: applies as-is → keep
• Partially relevant: valid insight but needs adaptation → recontextualize
• Irrelevant: doesn't apply → dismiss with reason

In --all-fresh mode, apply the relevance check to ALL arguments from all agents.

Step 2: Lone wolf insights

Highlight arguments from the lone wolf that survived cross-critique. These are the high-value outsider perspectives — things insiders missed because of proximity to the project. Flag them prominently in the Debate Record.

Step 3: Synthesis

Produce the Debate Record. Be ruthlessly honest:

• An argument that survived cross-critique AND relevance check IS strong — don't downplay it
• An argument that was refuted IS weak — don't rescue it
• A tension that persists IS real — don't pretend consensus exists
• An argument dismissed for irrelevance must state WHY it doesn't apply here

Debate Record format

Write to .meta/debates/debate-{slug}.md:

# Debate Record — {subject}

**Date:** {date}
**Preset:** {preset}
**Mode:** {standard|deep}
**Context:** {hybrid|all-fresh}
**Wolf:** Agent C ({persona name})
**Status:** USER DECISION NEEDED

---

## Subject

{description of the debate topic and why it matters}

## Angles

- **A ({name}):** {one-line perspective} [insider]
- **B ({name}):** {one-line perspective} [insider]
- **C ({name}):** {one-line perspective} [lone wolf]

## Strong arguments (survived cross-critique)

- {argument} — raised by {agent}, uncontested / defended against {counter}
- ...

## Lone wolf insights

- {argument from wolf that survived} — why it matters: {insight}
- ...

## Contested arguments (no consensus)

- {argument} — {agent} for, {agent} against
  - **For:** {reasoning}
  - **Against:** {reasoning}
  - **Why it matters:** {what changes depending on who's right}
- ...

## Dismissed arguments (refuted or irrelevant)

- {argument} — raised by {agent}, refuted by {agent}: {reason}
- {argument} — raised by {agent}, dismissed: irrelevant to project because {reason}
- ...

## Convergence points

- {point all 3 agents agree on}
- ...

## Irreducible tensions

- {tension that cannot be resolved — requires a value judgment}
- ...

## Recommendation

{orchestrator's synthesis — not a decision, a reasoned recommendation
based on argument strength. State confidence level.}

## Decision

**USER DECISION NEEDED**

To decide, consider:
- {key question 1 that tips the balance}
- {key question 2}

Domain presets

Each preset defines default angles and a natural lone wolf (3rd persona). The orchestrator may reassign the wolf if the subject demands a different outsider perspective.

default

Angles: Advocate / Critic / Pragmatist (wolf)

Attack taxonomy:

• Confirmation bias: are you only seeing evidence that supports your view?
• Survivorship bias: are you ignoring the failures?
• False dichotomy: is there a third option nobody considered?
• Scope creep: does this solve the actual problem or a bigger one nobody asked for?
• Status quo bias: are you defending the current approach just because it's current?

architecture

Angles: Puriste (correctness, patterns) / Pragmatique (ship fast, iterate) / End-user (simplicity, DX) (wolf)

Attack taxonomy:

• Over-engineering: building for hypothetical scale that may never come
• Premature abstraction: abstracting before understanding the concrete cases
• Resume-driven development: choosing tech for CV, not for the problem
• YAGNI violation: adding flexibility nobody asked for
• Scaling fallacy: "but what if we have 10M users" when you have 10
• Consistency trap: forcing consistency where context differs

strategy

Angles: Optimiste (opportunity, growth) / Sceptique (risk, cost) / Outsider (competitive landscape) (wolf)

Attack taxonomy:

• Survivorship bias: only looking at companies that succeeded with this approach
• Sunk cost: continuing because of past investment, not future value
• Market timing: assuming conditions will remain favorable
• Competitive blind spot: ignoring what others are doing differently
• First-mover fallacy: assuming first = winner
• Narrative bias: a compelling story is not evidence

data

Angles: Statisticien (rigor, methodology) / Ingénieur (performance, scalability) / Business (ROI, actionability) (wolf)

Attack taxonomy:

• Data leakage: using future information to predict the past
• Sampling bias: your data is not representative
• Feature engineering trap: complexity that doesn't improve signal
• Metric gaming: optimizing the metric, not the goal
• Cost/value ratio: is the accuracy gain worth the compute?
• Reproducibility: can someone else get the same results?

academic

Angles: Formaliste (rigor, proof) / Empiriste (evidence, experiments) / Reviewer (clarity, contribution) (wolf)

Attack taxonomy (inspired by Math Olympiad adversarial prompts):

• Tautological reduction: the argument is circular
• Proves too much: the reasoning would prove something known to be false
• Specification gaming: answering an easier version of the question
• Hand-wave at the crux: "by standard methods" at the hardest step
• Cherry-picked evidence: ignoring contradictory results
• Novelty inflation: repackaging known results as new

security

Angles: Attacker (exploit perspective) / Defender (mitigation, cost) / Compliance (regulation, audit) (wolf: Attacker)

Note: For security, the Attacker is the natural wolf — the outsider who thinks like a threat actor, not like the team defending the system.

Attack taxonomy:

• Security theater: measures that look good but don't help
• False sense of security: one strong lock on a door with no walls
• Threat model gap: protecting against the wrong attacker
• Usability trade-off: security so strict nobody follows it
• Compliance checkbox: meeting the letter, not the spirit
• Cost asymmetry: defense cost vs attack cost

Implementation notes

• Use the Agent tool with run_in_background: false for Phase 1 and Phase 2 (we need results before proceeding to the next phase)
• Insiders: use model: "haiku" to save cost and increase speed
• Lone wolf: use model: "sonnet" — higher capability compensates for less context
• The orchestrator (main agent) handles synthesis with the full model
• Create .meta/debates/ directory if it doesn't exist
• Slug format: debate-{kebab-case-subject}.md
• If a debate record with the same slug exists, append -2, -3, etc.