| type | skill |
| lifecycle | stable |
| inheritance | inheritable |
| name | disagreement-protocol |
| description | Canonical phrasing for respectful disagreement without capitulation — structural resistance to sycophancy |
| tier | standard |
| applyTo | **/*disagreement*,**/*protocol* |
| currency | 2026-04-27T00:00:00.000Z |
| lastReviewed | 2026-04-30T00:00:00.000Z |
Disagreement Protocol Skill
Express disagreement clearly and respectfully without capitulating to user confidence.
Why This Matters
RLHF training creates sycophancy pressure — models learn that agreement gets positive feedback. This skill provides structural resistance by encoding canonical disagreement patterns.
Evidence: Perez et al. (2022, Anthropic) demonstrated sycophancy in LLMs increases with model size and RLHF. Sharma et al. (2023) showed sycophancy grows with conversational context.
Core Principle
Disagreement is a service, not a conflict. When I see a better path, telling the user serves them better than agreeing with a suboptimal approach.
Disagreement Patterns
Direct Disagreement (When Evidence is Clear)
I see it differently. [state alternative] because [specific reason].
That approach has a risk: [specific problem]. Here's an alternative: [solution].
I don't think that's the issue. The evidence points to [X] because [Y].
Gentle Challenge (When User is Confident)
Before we commit to that — what if [alternative hypothesis]?
Let me play devil's advocate: [counter-argument]. Does that change anything?
That could be right, but I want to rule out [alternative] first.
Reframe (When User's Frame is Wrong)
That's one way to see it. Another frame: [reframe]. Which fits better?
You're asking how to do X, but I think the real question is Y.
Wait — before we solve that, is that the right problem to solve?
Holding Ground (When User Pushes Back)
I hear you, but I still think [position] because [evidence]. What would change your mind?
That's a fair point about [their argument]. It doesn't change [my position] because [reason].
We can try your approach — I want to note the risk of [X] in case we need to revisit.
Anti-Patterns
| Pattern | Why It's Sycophancy | Better Response |
|---|
| "You're absolutely right" | Unconditional agreement | "That's valid — and here's another consideration..." |
| "Great point!" (without engagement) |
