| name | deep-review |
| description | Adversarial code review with three parallel perspectives — Advocate, Skeptic, Architect — that create productive tension. Use for high-stakes PRs, architectural changes, or when single-pass review would miss issues. Surfaces findings through disagreement, not consensus. |
| lastReviewed | 2026-07-28T00:00:00.000Z |
Deep Review
Perform thorough code review using three perspectives with opposing mindsets. Their disagreement surfaces issues; their agreement signals confidence.
When to Use
- Architectural changes, multi-file refactors, or security-sensitive code
- PRs that are too important for single-pass review
- When you suspect confirmation bias in a standard review
- High-stakes merges where the cost of a missed issue is high
When NOT to Use
- Routine single-file edits (use standard
code-review skill)
- Documentation-only PRs
- Formatting/linting changes
The Three Perspectives
| Agent | Mindset | Question | Owns |
|---|
| Advocate | "Why is this correct?" | Trust boundaries, design rationale, false-positive defense | Correctness defense |
| Skeptic | "How can I break this?" | Bugs, edge cases, code smells that indicate bugs | Correctness attack |
| Architect | "Is this the right direction?" | System impact, scope, structural smells, tech debt | Direction |
Workflow
Phase 1: Gather Context
- Identify the changes — PR diff, local changes, or specific files
- Collect context — related files, tests, recent history of changed modules
- Note observations — anything unusual before analysis begins
Phase 2: Parallel Analysis
Run all three perspectives independently. Each sees the same context but asks different questions.
Same-model role separation vs cross-model separation — this three-perspective discipline works on a single model because the roles separate the analysis, not the models. For high-stakes reviews where "same-model shares the proposer's blind spots" is a real concern, compose with the Cross-Model External Critic method in adversarial-review — run this three-perspective analysis with a different model on fresh context.
Advocate Analysis
- What problem does this solve?
- What design decisions are intentional (not accidental)?
- Where are the trust boundaries correctly placed?
- What would break if we rejected this PR?
- Defend against false-positive concerns raised by Skeptic
Skeptic Analysis
- What inputs could break this? (null, empty, overflow, concurrent, malicious)
- What error paths are unhandled?
- What assumptions are undocumented?
- What would a fuzzer find?
- What code smells indicate deeper bugs? (naming lies, magic numbers, commented-out code)
- What works in tests but would fail in production?
Architect Analysis
- Does this fit the existing architecture or fight it?
- What's the blast radius if this fails?
- Does this increase or decrease coupling?
- Is there scope creep disguised as "while I'm here"?
- What precedent does this set for future changes?
- Is there tech debt being introduced? Is it intentional and documented?
Phase 3: Synthesis
3.1 Agreement Analysis
What do multiple perspectives agree on? → High-confidence findings.
3.2 Conflict Resolution
When perspectives disagree, apply these rules:
| Conflict | Resolution |
|---|
| Skeptic finds bug, Advocate defends | Does Advocate cite file:line that refutes? If not, Skeptic wins |
| Advocate says intentional, Skeptic says bug | If Skeptic shows reproducible path → it's a bug regardless of intent |
| Architect says blocking, Skeptic disagrees on priority | Skeptic's priority on correctness issues; Architect's on direction |
| No evidence either way | Mark as "Disputed" for human decision |
Core rule: Evidence beats assertion. A file:line citation wins over "probably."
3.3 Final Output
## Deep Review: <title>
### Summary
<1-2 sentence overview of the change and verdict>
### Perspectives
**Advocate** (Design Rationale)
<key defenses and intentional design decisions>
**Skeptic** (Risk Analysis)
<bugs found, edge cases, concerns with evidence>
**Architect** (Architectural Impact)
<patterns, debt, direction, system-level concerns>
### Consolidated Findings
| # | Issue | Priority | Advocate | Skeptic | Architect |
|---|-------|----------|----------|---------|-----------|
| 1 | <issue> | Critical/High/Medium/Low | <view> | <view> | <view> |
### Disputed (if any)
<issues where perspectives disagree and human must decide>
### Recommendations
<prioritized actions>
### Follow-up Items
Priority Classification
| Priority | Criteria | Action |
|---|
| Critical | Data loss, security vulnerability, crash in production path | Block merge |
| High | Incorrect behavior under realistic conditions | Block merge |
| Medium | Code smell, missing test, unclear naming, minor debt | Request fix or accept with note |
| Low | Style, nitpick, suggestion for future | Comment only |
Example
PR: Add rate limiting middleware to API gateway
Advocate: "Rate limiting prevents resource exhaustion. The sliding window approach handles burst traffic better than fixed windows. The 429 response includes Retry-After header per RFC 6585."
Skeptic: "The window reset logic at middleware/rate-limit.ts:47 uses Date.now() but the TTL in Redis uses seconds — off by 1000x. Under load, the counter will never expire. Also: no test covers the window boundary."
Architect: "Rate limiting belongs at this layer (before auth, after TLS). But the config is hardcoded — should use env vars for per-deployment tuning. This sets precedent that all middleware reads config from constants."
Synthesis: Skeptic's timing bug is Critical (blocks merge). Architect's config concern is Medium (fix in follow-up). Advocate's design rationale is sound.
Integration with ACT
- Tenet II (Disconfirmation): The Skeptic's entire job is disconfirmation
- Tenet III (Multiple Hypotheses): Three perspectives prevent anchoring on first interpretation
- Tenet VIII (Adversarial Self-Probe): The review structure IS adversarial by design
- Materiality Gate: Use Deep Review for high-stakes; standard
code-review for routine
Would Revise If
Revise if the three-perspective (Advocate / Skeptic / Architect) review repeatedly converges to consensus that misses real defects later found in production, or if reviewer time-cost exceeds 2× standard code-review without producing distinct findings in 3+ consecutive uses.
