Dissent

/dissent

Structured disagreement that strengthens decisions. The insight: find flaws before the one-way door closes.

Dissent is not attack. It's the practice of actively seeking reasons you might be wrong. The devil's advocate is a role, not a personality.

When to Use

Invoke /dissent when:

• About to lock in a one-way door - architecture choices, major hires, public API contracts, anything hard to reverse
• Confidence is high but stakes are higher - feeling certain is when you need dissent most
• Team is converging too quickly - unanimous agreement without debate is a warning sign
• You're defending a position - advocacy mode is the enemy of truth-seeking
• The path forward seems obvious - obvious paths have hidden assumptions

Do not use when: Gathering initial options, brainstorming, or exploring. Dissent is for stress-testing decisions, not generating them.

The Dissent Process

Step 1: Steel-Man the Current Approach

Before attacking, fully articulate the position you're challenging:

"The current approach is [approach]. The reasoning is [reasoning]. The expected outcome is [outcome]. This is the strongest version of this position."

If you can't state the position charitably, you don't understand it well enough to challenge it.

Step 2: Seek Contrary Evidence

Actively search for information that contradicts the current approach:

• What data would prove this approach wrong?
• Who disagrees with this? What's their strongest argument?
• What similar approaches have failed elsewhere? Why?
• What are we ignoring because it's inconvenient?

"If this approach were wrong, what would we expect to see? Are we seeing any of that?"

Step 3: Pre-Mortem

Imagine it's six months from now and this decision failed. Work backward:

"This failed because [reason]. The warning signs we ignored were [signs]. The assumption that broke was [assumption]."

Generate at least three plausible failure scenarios:

1. Technical failure - It doesn't work as expected
2. Adoption failure - It works but nobody uses it / changes nothing
3. Opportunity cost - It works but we should have done something else

Step 4: Surface Hidden Assumptions

Every decision rests on assumptions. Most aren't stated. Find them:

• What are we assuming about the user/customer?
• What are we assuming about the system/codebase?
• What are we assuming about the timeline/resources?
• What are we assuming won't change?

For each assumption:

Assumption: [what we're taking for granted]
Evidence: [what supports this assumption]
Risk if wrong: [what happens if this assumption breaks]
Test: [how we could validate this before committing]

Step 5: Decide

Based on the dissent analysis, recommend one of:

• PROCEED - Dissent found no critical flaws. Decision strengthened.
• ADJUST - Dissent surfaced issues that can be addressed. Modify the approach.
• RECONSIDER - Dissent revealed fundamental problems. Go back to solution space.

Always include the reasoning:

"PROCEED: The strongest counter-argument is [X], but it's addressed by [Y]. The key assumption is [Z], which we've validated by [how]."

ADR Generation

If the decision is a one-way door (hard to reverse) and you recommend PROCEED or ADJUST, offer to create an Architecture Decision Record (ADR). Future team members will ask "why did we do it this way?"

Offer to create an ADR when:

• The decision affects system architecture
• Multiple valid approaches were considered and rejected
• The reasoning depends on current context that might not be obvious later
• Someone will likely question this in 6 months

The dissent report maps directly to ADR format:

If user accepts, write to: docs/adr/NNNN-<decision-slug>.md (or project's ADR location)

# ADR NNNN: [Decision Title]

## Status
Accepted

## Context
[From Steel-Man Position - why this decision needed to be made]

## Decision
[From Decision section - what we decided and why]

## Options Considered
[From Contrary Evidence - what alternatives were evaluated]

## Consequences
[From Hidden Assumptions - what we're accepting by making this decision]

## Notes
Generated from /dissent on [date]

Output Format

Always produce a dissent report in this structure:

## Dissent Report

**Decision under review:** [what's being challenged]
**Stakes:** [why this matters]
**Confidence before dissent:** [HIGH/MEDIUM/LOW]

### Steel-Man Position
[The strongest version of the current approach]

### Contrary Evidence
1. [Evidence point 1]
2. [Evidence point 2]
3. [Evidence point 3]

### Pre-Mortem Scenarios
1. **[Failure mode]:** [How this could fail]
2. **[Failure mode]:** [How this could fail]
3. **[Failure mode]:** [How this could fail]

### Hidden Assumptions
| Assumption | Evidence | Risk if Wrong | Test |
|------------|----------|---------------|------|
| [Assumption 1] | [Evidence] | [Risk] | [Test] |
| [Assumption 2] | [Evidence] | [Risk] | [Test] |

### Decision

**Recommendation:** [PROCEED / ADJUST / RECONSIDER]

**Reasoning:** [Why this recommendation]

**If ADJUST:** [Specific modifications]

**Confidence after dissent:** [HIGH/MEDIUM/LOW]

**Create ADR?** [Yes - offer to write docs/adr/NNNN-slug.md / No - not a one-way door]

Examples

Example 1: Architecture Decision

/dissent

We're about to adopt a microservices architecture for the new
payments system. The team agrees this is the right approach.

Output:

## Dissent Report

**Decision under review:** Microservices architecture for payments
**Stakes:** 6+ month commitment, affects entire system
**Confidence before dissent:** HIGH (team unanimous)

### Steel-Man Position
Microservices will allow independent scaling of payment processing,
enable different teams to deploy independently, and provide fault
isolation so one service failure doesn't bring down payments.

### Contrary Evidence
1. We have 4 engineers. Netflix had 100+ when they went microservices.
2. Our deployment pipeline takes 45 minutes. We'll multiply that by N services.
3. The "independent teams" benefit requires teams we don't have.

### Pre-Mortem Scenarios
1. **Distributed systems complexity:** We spend 60% of time debugging
   cross-service issues instead of building features.
2. **Operational overhead:** We can't hire SREs fast enough. System
   is down because nobody understands the full architecture.
3. **Premature optimization:** The "scaling" benefit never materializes
   because we never got enough users to need it.

### Hidden Assumptions
| Assumption | Evidence | Risk if Wrong | Test |
|------------|----------|---------------|------|
| We'll need to scale components independently | None - projected | Waste of complexity | Check actual load patterns |
| Team will grow to support this | Hiring plan exists | Stuck with complexity, no staff | Validate funding/timeline |
| Network latency acceptable | Not measured | Payment failures | Prototype and benchmark |

### Decision

**Recommendation:** RECONSIDER

**Reasoning:** The benefits of microservices assume scale and team
size we don't have. The strongest argument for microservices (independent
scaling) has no evidence supporting the need. A modular monolith gives
us the code organization benefits without the operational complexity.

**Confidence after dissent:** HIGH (clear mismatch)

**Create ADR?** No - we're reconsidering, not committing

Example 2: Feature Decision

/dissent

We're about to ship a recommendation system that suggests similar
items based on purchase history. Built over 3 sprints.

Output:

## Dissent Report

**Decision under review:** Ship purchase-history recommendation system
**Stakes:** 3 sprints invested, prominent homepage placement
**Confidence before dissent:** MEDIUM

### Steel-Man Position
Users who bought X often want Y. Amazon proved this works. We have
the purchase data. The model is trained and performs well in testing.

### Contrary Evidence
1. Our testing used historical data. Users might behave differently
   when seeing recommendations in real-time.
2. We have 10K users. Amazon's model works with billions of data points.
3. Competitor tried this, removed it after 6 months (no lift).

### Pre-Mortem Scenarios
1. **Cold start problem:** New users see garbage recommendations,
   leave before making a purchase we could learn from.
2. **Filter bubble:** System recommends what users already buy,
   doesn't drive discovery of new product categories.
3. **Wrong metric:** Recommendations get clicks but not purchases.
   We optimized for engagement, needed revenue.

### Hidden Assumptions
| Assumption | Evidence | Risk if Wrong | Test |
|------------|----------|---------------|------|
| Purchase history predicts intent | Industry standard | Wasted real estate | A/B test against simple "popular items" |
| 10K users is enough data | None | Poor recommendations | Check minimum viable data in literature |
| Homepage placement is right | None | Users ignore it | Heatmap existing traffic patterns |

### Decision

**Recommendation:** ADJUST

**Reasoning:** The approach is sound but we're shipping without
validation. Risk is manageable with modifications.

**Modifications:**
1. Ship with A/B test against "popular items" baseline
2. Add fallback to curated picks for users with <5 purchases
3. Define success metric (revenue lift, not clicks) before launch

**Confidence after dissent:** MEDIUM (reduced risk with A/B)

**Create ADR?** Yes - shall I write `docs/adr/0012-recommendation-system-rollout.md`? Documents why we chose A/B testing and what success metrics we committed to.

Session Persistence

This skill can persist context to .oh/<session>.md for use by subsequent skills.

If session name provided (/dissent auth-decision):

• Reads/writes .oh/auth-decision.md directly

If no session name provided (/dissent):

• After producing the dissent report, offer to save it:

  "Save to session? [suggested-name] [custom] [skip]"

• Suggest a name based on git branch or the decision topic

Reading: Check for existing session file. Read Aim, Problem Statement, Solution Space to understand the decision being challenged.

Writing: After producing the dissent report:

---
session: <session>
artifact_type: dissent
updated: <timestamp>
---
## Dissent
**Updated:** <timestamp>
**Decision:** [PROCEED | ADJUST | RECONSIDER]

[dissent report content]

Adaptive Enhancement

Base Skill (prompt only)

Works anywhere. Produces dissent report for manual review. No persistence.

With .oh/ session file

• Reads .oh/<session>.md for context on the decision
• Writes dissent report to the session file
• Subsequent skills see that dissent was performed

With Open Horizons MCP

• Queries past decisions on similar topics
• Retrieves relevant guardrails and tribal knowledge
• Logs dissent decision for future reference
• Session file serves as local cache

With RNA MCP (repo-native-alignment)

• Call search("risks and constraints for [area]", include_artifacts: true) to ground dissent
• Call outcome_progress to assess whether the approach serves the outcome
• Present dissent findings as candidate learnings and request explicit human selection.
• Write only human-approved candidates to .oh/metis/; compact or dismiss the rest.

With Team Input

• Aggregates dissent from multiple reviewers
• Tracks which assumptions different team members question

Position in Framework

Combines with: /solution-space (challenge the recommendation), /problem-statement (challenge the framing), /execute (before one-way doors). Leads to: PROCEED (continue with confidence), ADJUST (modify approach), or RECONSIDER (back to solution-space). This is not a phase: Dissent is an overlay you invoke when stakes are high.

Leads To

After dissent, typically:

• PROCEED - Continue to /execute or /ship with strengthened confidence
• ADJUST - Update the approach, then /review the modifications
• RECONSIDER - Return to /solution-space with new constraints

Remember: Dissent is not doubt. It's the discipline of seeking truth before comfort. The goal isn't to stop decisions—it's to make better ones.