| name | multi-ai-review |
| description | Get a second (and third) AI opinion on hard-to-reverse decisions. Runs the same prompt through different agent roles or model families in parallel, then reconciles disagreements. Inspired by gstack's `/codex` cross-review pattern. Activated for architecture decisions, security-sensitive code, pricing/packaging changes, deploy gates, legal-adjacent text, or any decision the user flags as "não dá pra desfazer fácil". |
Multi-AI Review
Solo reasoning ships solo blind spots. For decisions with hard-to-reverse blast radius, get at least one independent reviewer with a different prior and explicitly reconcile disagreements before committing.
This is not ceremony. It is the cheapest mistake-prevention you can buy in an AI-assisted workflow: minutes of compute against hours-to-days of cleanup.
When to invoke
Mandatory:
- Architecture decisions touching auth, data model, or payments.
- Migrations, schema changes, or anything irreversible without a backup restore.
- Pricing or packaging changes (also triggers
business-plan-impact-review).
- Public-facing legal/compliance text (Privacy Policy, Terms, refund policy, LGPD/GDPR statements).
- Any deploy gate where rollback is non-trivial.
- Adoption of a new external skill bundle (
#agents-marketplace packs).
- The user explicitly says "isso é importante" / "não posso errar aqui" / "double-check isso".
Optional but recommended:
- New feature spec right before sprint commitment.
- Naming a public-facing thing you can't easily rename later (product name, API endpoint, slug).
- Choosing between two non-trivially-different stacks.
Skip:
- Routine code edits, refactors, bug fixes with tests.
- Reversible UI tweaks.
- Internal-only docs.
Reviewer roles (pick at least 2)
The skill orchestrates multiple reviewers, each with a distinct lens. The user picks the relevant 2–4 per decision. Default set in bold.
| Role | Lens | Existing agent (if any) |
|---|
| Devil's advocate | "Why is this wrong?" | .claude/agents/devils-advocate-agent.md |
| Technical/security red team | "How does this break or get exploited?" | .claude/agents/technical-security-red-team-agent.md |
| Business red team | "Why does this not make business sense?" | .claude/agents/business-red-team-agent.md |
| Privacy/compliance | "Where does this leak personal data or trip LGPD/GDPR?" | uses privacy-audit skill |
| Cost watchdog | "Where does this silently get expensive?" | uses cost-watchdog skill |
| Plain-Portuguese reader | "Would a non-dev customer understand this?" | uses plain-portuguese-explainer skill |
| Independent model | Same prompt, different model family (e.g. Sonnet ↔ Opus, or external /codex) | external |
Always include at least one reviewer whose model family OR agent role is different from the agent that produced the artifact. Same-model self-review is weaker than independent review.
Procedure
- Frame the decision artifact. Provide the reviewers with: the decision/spec/diff, the explicit blast-radius statement, and the alternatives considered.
- Dispatch reviewers in parallel. Each reviewer answers the same 4 questions:
- What is your single biggest concern?
- What is the most likely failure mode?
- What is one alternative that should also be considered?
- On a 0–10 scale, how confident are you the proposed direction is correct?
- Collect raw outputs verbatim. Do not pre-summarize.
- Reconcile disagreements. Produce a reconciliation note that lists:
- Where reviewers agreed (treat as locked-in concerns).
- Where reviewers disagreed (require human judgment).
- Net confidence (lowest individual score, not the average).
- Decide. Three possible outcomes:
- Proceed — all reviewers ≥7 confidence, no unresolved high-severity disagreement.
- Modify and re-review — one or more concerns require artifact change; re-run with the modified artifact.
- Block — at least one reviewer ≤4 confidence on a hard-to-reverse aspect. Decision needs more research, alternatives, or human deliberation.
- Log in
session-log/. Always log the reconciliation, even (especially) for "Proceed" — future-you needs to know the decision was reviewed.
Output format
## Multi-AI Review — <decision name>
**Date:** <YYYY-MM-DD>
**Decision artifact:** <link to spec / diff / file>
**Blast radius:** <reversible / cosmetic / financial / customer-trust / P0>
**Reviewers:** <list of roles>
### Raw reviewer outputs
#### <Reviewer 1 role>
- Biggest concern: ...
- Most likely failure mode: ...
- Alternative: ...
- Confidence: <0–10>
#### <Reviewer 2 role>
...
### Reconciliation
- **Agreed concerns:** ...
- **Disagreements requiring human judgment:** ...
- **Net confidence (min):** <0–10>
### Decision
- [ ] Proceed
- [ ] Modify and re-review (next iteration: <YYYY-MM-DD>)
- [ ] Block
### Action items
- ...
### Cross-link
- Session log entry: `session-log/<date>-<slug>.md`
Anti-patterns (will fail the review)
- Same-model echo chamber. Asking the same Claude-Opus instance "are you sure?" is not multi-AI review.
- Average-the-confidence. Net confidence is the minimum of individual scores, not the mean. The most worried reviewer represents the least-explored failure mode.
- Ignore the dissent. A reviewer scoring 4 with a specific concern is more valuable than three reviewers scoring 9 with vague approval.
- Run after committing. Review BEFORE the irreversible action. Post-hoc review is a session log, not a gate.
- Skip the log. A review that wasn't logged didn't happen, from the perspective of the next session.
Cost guardrails
Each reviewer is a separate model invocation. For cost-conscious projects:
- Default to 2 reviewers, escalate to 4 only for P0 blast radius.
- Prefer existing in-repo agents (
devils-advocate-agent, technical-security-red-team-agent, business-red-team-agent) over spawning external models for routine cases.
- Use
cost-watchdog to verify the cumulative review cost is within the project's monthly LLM budget, especially if the skill is invoked >5x/week.
Cross-references