| name | adversarial-review |
| type | review |
| enforcement | invocable |
| priority | high |
| description | Sends a commit range (default: origin/main..HEAD) to an external LLM
with the framing "find every little thing wrong as though you're trying
to take the engineer's job." Returns a structured findings report under
audit-reports/. Designed as the external-audit backstop named in
careful-not-clever v1.8.1-alpha §"The limit of this rule" — the rule
acknowledges that an author cannot reliably catch their own claim-
evidence gaps, and points at an external pass as the real check.
|
| activates_on | {"slash_commands":["/adversarial-review","/adversarial-review <base-ref>","/adversarial-review <base-ref> <model>"],"keywords":["external audit","adversarial review","send to grok","second opinion","review my commits","audit this branch"],"intent_patterns":["ask <model> to review","external pass","second pair of eyes"]} |
| implementation | admin/external-audit.sh |
Adversarial Review
Why this skill exists
Careful-not-clever v1.8.1-alpha §"The limit of this rule" acknowledges that the claim-evidence table is a forcing function, not a guarantee. Two failure modes survive even with the table in place:
- Misuse mode A — Vague evidence: "tests pass" / "verified" with no specific artifact
- Misuse mode B — Narrow claim: the claim is technically supported but is narrower than the actual scope of the change
Mode A a careful reviewer can spot. Mode B is harder for the author to see than vague evidence, because the entry passes their own self-review (the evidence really does support what they wrote). The rule says: "The most reliable check is external."
This skill is that external check, automated. It is a triage aid, not a gate.
When to invoke
| Situation | Invoke |
|---|
| Before a PR / merge to main on any Layer 2 or Layer 3 change | YES |
| After completing a multi-commit batch where you suspect blind spots | YES |
| Before declaring a long-running thread "complete" | YES |
| Routine refactor with full test coverage | OPTIONAL |
| Trivial typo fix | NO (wasteful) |
Cost per invocation (Grok): roughly $0.04 for a 9-commit batch (in=5555, out=1768 tokens at grok-3 pricing). Smaller ranges cost proportionally less. GPT-4o on the same range runs about $0.02 but produces fewer unique findings — see "Choosing a model" below.
How to invoke
Slash command
/adversarial-review # audits origin/main..HEAD with grok
/adversarial-review HEAD~3 # audits the last 3 commits with grok
/adversarial-review origin/main gemini # different reviewer
Direct script
admin/external-audit.sh
admin/external-audit.sh HEAD~5 grok
admin/external-audit.sh origin/main gpt
Choosing a model
Empirical from two reviews (audit-reports 2026-05-13):
| Model | Cost / 9 commits | Findings produced |
|---|
