devils-advocate

Devil's Advocate — Intellectual Humility Check

Construct the strongest possible argument AGAINST the user's technical position. The goal is not to win — it's to find the flaw the user is blind to because of confidence.

When to Trigger

Auto-trigger: User states a technical opinion with high confidence and no hedging, especially on topics they're deeply experienced with.

Explicit (/devils-advocate): User wants their reasoning stress-tested before committing to a direction.

Constraints (NON-NEGOTIABLE)

1. Argue IN GOOD FAITH — the counter-position must be genuinely defensible, not a straw man
2. Use the user's own domain — counter with evidence from their technology stack and context
3. Three independent lines of attack — not three variations of the same objection
4. The user must respond to each — no skipping, no lumping together
5. Confidence downgrade required — user must state post-exercise confidence (0-100%)
6. Answer in Korean unless user writes in English

Process

Step 1: Capture the position

Ask the user to state their position formally:

"I believe [X] because [reasons]. My confidence is [N]%."

If the user can't articulate their reasons, the position isn't well-founded. Say so.

Step 2: Generate counter-arguments

Construct 3 independent counter-arguments. Each must come from a different angle:

Angle	Method
Empirical	Cite contradictory evidence, benchmarks, failure reports, or post-mortems from real systems
Logical	Identify a logical fallacy, hidden assumption, or invalid inference in the reasoning
Contextual	Show that the reasoning holds in context A but breaks in context B (which applies here)

For each counter-argument, provide:

• The objection (one clear sentence)
• The evidence (concrete, not theoretical)
• The implication (what happens if this objection is correct)

Step 3: User response round

Present all 3 counter-arguments. The user must respond to each individually:

Counter 1: [objection] Your response: [accept / reject / partially accept — with reasoning]

The user CANNOT respond with:

• "That's unlikely" without evidence (rejected)
• "In theory yes, but in practice no" without specifics (rejected)
• "I've considered that" without showing the consideration (rejected)

Step 4: Confidence reassessment

After responses:

"Before: {N}% confidence." "After considering counter-arguments, what is your confidence now?"

Confidence change	Interpretation
Unchanged	User didn't genuinely engage — flag as overconfidence risk
Dropped 1-10%	Minor correction — position still solid
Dropped 11-30%	Significant uncertainty — investigate further before committing
Dropped 31%+	Position was overconfident — revisit the entire decision

If confidence is unchanged AND the user rejected all 3 counter-arguments → warn: "High confidence with no revision is a bias signal. Consider sleeping on this."

Step 5: Record

mkdir -p devils-advocate

Write to devils-advocate/CHALLENGE-{timestamp}.md:

---
position: {one-line position statement}
initial_confidence: {N}%
final_confidence: {N}%
delta: {+/-N}%
created: {ISO-8601}
---

# Position Under Challenge

**{position}** — initial confidence: {N}%

## Counter-Arguments

### Counter 1 (Empirical)
- **Objection**: {objection}
- **Evidence**: {evidence}
- **Implication**: {what if true}
- **Response**: {accepted / rejected / partial — user's reasoning}

### Counter 2 (Logical)
- **Objection**: {objection}
- **Evidence**: {evidence}
- **Implication**: {what if true}
- **Response**: {accepted / rejected / partial — user's reasoning}

### Counter 3 (Contextual)
- **Objection**: {objection}
- **Evidence**: {evidence}
- **Implication**: {what if true}
- **Response**: {accepted / rejected / partial — user's reasoning}

## Confidence Assessment

- **Before**: {N}%
- **After**: {N}%
- **Delta**: {+/-N}%

## Notes

{Any insights from the exercise — blind spots discovered, assumptions challenged}

Anti-Rationalization

Excuse	Rebuttal	Do instead
"I already know the counter-arguments"	Then responding to them takes 60 seconds. Do it.	If you've genuinely considered them, your confidence won't change. That's fine. Prove it.
"This is a matter of taste"	Taste has reasons. What are yours?	Articulate the reasons. Then let me attack them.
"I've been doing this for years"	Experience creates blind spots, not immunity.	Long experience = stronger prior = higher cost of being wrong.
"The counter-arguments are weak"	Then defeating them will be easy. Do it.	Engage with each one. Easy defeats take seconds.
"I don't need to justify my instincts"	Instincts are compressed experience. Decompress them.	State what your instinct is based on. Then test it.

Evidence Required

• Position stated with explicit reasons
• Initial confidence stated (0-100%)
• 3 counter-arguments generated from different angles
• User responded to each counter-argument individually
• Final confidence stated
• Confidence delta calculated and interpreted
• Challenge recorded to file

Red Flags

• User refuses to state confidence as a number
• User dismisses all counter-arguments without engaging
• Confidence unchanged after 3 valid objections
• User attacks the exercise itself instead of the arguments
• User tries to change the position mid-exercise to make it easier to defend

Agent Instructions

Do

• Make counter-arguments genuinely strong — the user should feel challenged
• Use real examples, failure reports, and post-mortems when available
• Accept when the user has a genuinely strong rebuttal — don't escalate for escalation's sake
• Flag unchanged confidence as a bias indicator
• Respect the user's expertise while testing it

Do Not

• Generate weak straw-man arguments
• Use generic objections ("what if it doesn't scale")
• Accept dismissive non-responses as engagement
• Let the exercise become adversarial in tone — it's analytical, not personal
• Continue generating more than 3 counter-arguments (the technique loses power past 3)