Adversarial Review

Adversarial Review

When to use

Use this skill when:

• You've completed a plan, design, or proposed fix and are about to present it.
• The change is non-trivial (affects multiple files, changes behavior, touches critical paths).
• You're about to recommend an architecture or design decision.
• The user submits finished work (draft, post, naming decision, design proposal) and asks for an honest critical take — "what do you actually think?", "be brutal", "was hältst du wirklich davon". The flow is the same Attack-Defend-Revise loop, but on the user's artifact rather than the agent's plan.

Do NOT use when:

• The task is trivial (renaming, formatting, simple config change).
• The user explicitly asked for a quick/rough draft.
• You're exploring options, not committing to one yet.

Inbound delegation: decision-record § Weighted-matrix mode delegates a scoped attack here after the weighted sums ("attack the winner, using the losing options' strongest criteria") — treat the matrix + sensitivity block as the artifact under review.

Procedure: Adversarial review

1. Inspect the artifact — Read the plan, diff, or draft you are about to critique; note its scope, assumptions, and the explicit asks before attacking.
2. Attack — Run Step 1 below as the grumpy senior engineer.
3. Defend — Run Step 2 as the balanced engineer; classify each criticism as must-fix / defer / reject.
4. Revise — Run Step 3 to fold valid fixes back in, then report every criticism with its disposition. Deciding which trade-offs matter is the user's pass, not this step's.

Step 1: Attack (Grumpy Senior Engineer)

Assume your plan/fix is flawed. Ask yourself:

• What's the weakest assumption?
• Where will this break under load, at scale, or with edge cases?
• What did I ignore or hand-wave?
• Is this over-engineered for the actual problem?
• Would a simpler approach work just as well?
• What will the next developer curse me for?

Step 2: Defend (Balanced Engineer)

Counter the criticism fairly:

• Which criticisms are valid and must be addressed now?
• Which are theoretical and can be deferred?
• What's the pragmatic middle ground?

Step 3: Revise

• Fix the valid issues in your plan/fix.
• Move deferred concerns to "Open Questions" or "Known Limitations".
• Present the improved version to the user.

Debate internally, report completely. The user sees the improved result rather than the blow-by-blow — but every criticism raised reaches the report with its disposition (must-fix / deferred / rejected, each with one line of reason). Do not decide on the user's behalf which concerns were worth their attention.

This clause used to read "surface only the trade-offs that need the user's input", and that is the pre-filter defect: an instruction to report a subset is followed literally — the review finds the defects, then withholds the ones it judged unimportant, and the loss is invisible because the withheld set leaves no trace. It matters here more than anywhere: self_review_gate.ts loads this file as the system prompt for the package's own CI self-review, so a suppression clause here suppresses findings on every pull request.

Context-specific attack questions

Feature plans / Architecture

• Is this the simplest solution that works?
• What happens when requirements change (and they will)?
• Are there hidden dependencies or coupling?
• Does this respect existing patterns or introduce a new one unnecessarily?
• What's the migration/rollback story?

Bug fixes

• Is this the root cause or just a symptom?
• Will this fix break something else?
• Does the fix handle the edge case that caused the bug?
• Is there a regression test that proves the fix works?
• Are there other places with the same bug pattern?

Code changes / Refactoring

• Would I understand this code in 6 months without context?
• Did I check all callers and downstream effects?
• Are the tests actually testing the right behavior?
• Did I introduce a new pattern where an existing one would work?

Database migrations

• Can this destroy or corrupt data?
• Is rollback possible?
• What happens to running queries during migration?
• Did I check the table size (large table ALTER can lock)?

API design

• Is this a breaking change?
• Is it consistent with existing endpoints?
• Are error responses clear and actionable?
• Did I consider pagination, filtering, versioning?

Security-sensitive changes

• Where is the attack surface I'm not seeing?
• Am I trusting user input anywhere?
• Are there authorization gaps?
• Would this pass a security review?

Integration with other skills

• feature-planning — adversarial review after Understanding Lock, before presenting the plan.
• bug-analyzer — review the proposed fix before implementing.
• code-review — self-review before creating a PR.
• laravel-migration (or framework-native equivalent) — review migration for data safety.
• api-design — review API design for consistency and breaking changes.
• security — review security-sensitive changes for attack surface.

RDP: fresh-context verifier as the default gate (structural)

Within the Reasoning Discipline Protocol the fresh-context verifier subagent is the default final gate — but, because it is a full extra inference pass, it fires only on the structural-complexity signal: ≥ 2 of {branching/conditional logic, ≥ 3 explicit must/must-not constraints, stateful operations, irreversibility} and estimated work ≥ ~1k tokens. Token length alone never triggers it. See rdp-gate (L12).

Validate

• Confirm each identified risk has a concrete mitigation or explicit acceptance.
• Verify the review produced at least one actionable finding (or explicit "no issues found").
• Check that the review did not just restate the plan — it must challenge assumptions.

Output format

1. Improved plan/code incorporating adversarial findings
2. Every criticism raised, each with severity and disposition (must-fix / deferred / rejected) and one line of reason. Not "top concerns" — a summary that keeps only the highlights is the pre-filter defect wearing an editor's hat.
3. Remaining open risks (if any) with severity rating

Gotcha

• Don't use this on trivial changes — it adds overhead without value on simple renames or config tweaks.
• The model tends to invent risks that don't exist. Ground every concern in actual code, not hypotheticals.
• Don't challenge the user's explicit requirements — challenge the implementation, not the goal.

Do NOT

• Do NOT present the raw adversarial debate to the user — only the improved result.
• Do NOT use this as an excuse to delay work — the review should take seconds, not minutes.
• Do NOT apply this to trivial changes — it adds overhead without value.
• Do NOT let the "grumpy engineer" kill good ideas — the balanced engineer must counter.
• Do NOT skip Step 3 (Revise) — attacking without improving is just complaining.

References

• Tree-of-Thoughts (ToT) — arxiv.org/abs/2305.10601 Deliberate problem-solving by exploring multiple reasoning branches. This skill adapts ToT by pitting a grumpy engineer against a balanced engineer — the branching happens between roles, not between thought-tree nodes.