一键导入
adversarial-review
Two independent reviewers must both approve high-stakes output before it ships, breaking the single-reviewer blind-spot problem.
用 Codex 或 Claude 帮你安装 复制这段 Prompt,粘贴到 Codex、Claude 或其他助手里,让它检查 Skill 页面并帮你完成安装。
菜单
Two independent reviewers must both approve high-stakes output before it ships, breaking the single-reviewer blind-spot problem.
用 Codex 或 Claude 帮你安装 复制这段 Prompt,粘贴到 Codex、Claude 或其他助手里,让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Build Flutter APK/IPA, generate changelog, update buildlog, and commit version bump. Handles platform selection, build mode, and artifact renaming.
Build Unity APK/IPA on macOS with Unity 6.x — pre-flight checks, version bump, iOS signing params, post-build verification, buildlog, and commit.
Evidence-driven bug investigation and fix pipeline. Use when something is broken, not working, or the root cause is ambiguous.
Team-orchestrated feature development with parallel specialist agents. Use when building a feature that benefits from multiple agents working concurrently. You act as Product Director coordinating the team.
Interactive agent picker for composing and dispatching parallel teams. Browse available agents across domains, pick a roster, dispatch in parallel.
ASO optimization pipeline for an app store listing. Use when optimizing an App Store or Play Store listing.
| name | adversarial-review |
| description | Two independent reviewers must both approve high-stakes output before it ships, breaking the single-reviewer blind-spot problem. |
| when_to_use | Use before shipping output where a missed defect is costly — production code without human review, customer-facing copy, claims/stats/API references, compliance-sensitive text, or batch-generated content. Also when the user says "make sure this is right", "double-check this", "review this twice", "is this safe to ship", or "adversarial review". Skip for internal drafts, exploration, and anything a build/test/lint already verifies deterministically — use `verification-loop` for those. |
| allowed-tools | Agent Read Grep Glob |
Source: adapted from the "Santa Method" in affaan-m/ecc
skills/santa-method(MIT), originally by Ronald Skelton. Wired to ourAgent()model+effort conventions (agents.md).
Core principle: a single agent reviewing its own output shares the biases and blind spots that produced the output. Two independent reviewers, with no shared context, both required to pass, break that failure mode. One reviewer catching an issue means the issue is real — the other's miss is exactly the blind spot this exists to eliminate.
Run deterministic checks first (verification-loop — build, typecheck, lint, tests). This
skill is for the semantic layer those can't catch: correctness of claims, hallucinated APIs,
completeness against the spec, compliance, internal consistency.
Launch both in a single message (parallel). For genuine independence, give them different
models or different lenses — identical reviewers share blind spots. Per agents.md, every
Agent() call sets model.
Agent(subagent_type: "nextc-ecc:code-reviewer", model: "sonnet",
description: "Adversarial reviewer A",
prompt: REVIEWER_PROMPT with lens="correctness + completeness")
Agent(subagent_type: "nextc-ecc:code-reviewer", model: "opus",
description: "Adversarial reviewer B",
prompt: REVIEWER_PROMPT with lens="security + edge cases + hallucinated references")
For non-code output (copy, docs, claims), use subagent_type: "claude" reviewers with the same
prompt shape.
Each reviewer receives the original spec AND the output, the rubric, and returns a structured verdict — not prose:
You are an INDEPENDENT reviewer. You have NOT seen any other review of this output.
Your job is to FIND PROBLEMS, not to approve. Default to FAIL when uncertain.
## Spec / requirements
{spec}
## Output under review
{output}
## Rubric (evaluate EACH criterion)
{rubric}
Return:
{ "verdict": "PASS" | "FAIL",
"checks": [ {"criterion": "...", "result": "PASS|FAIL", "detail": "exact problem"} ],
"critical_issues": ["blockers that must be fixed"],
"suggestions": ["non-blocking"] }
Invariants: context isolation (neither reviewer sees the other), identical rubric, same inputs, structured output.
Every criterion needs an objective pass/fail condition. Baseline set:
| Criterion | Pass condition | Failure signal |
|---|---|---|
| Factual accuracy | Claims verifiable against source/common knowledge | Invented stats, wrong versions, nonexistent APIs |
| Hallucination-free | No fabricated entities, quotes, URLs, references | Links to nonexistent pages, unsourced quotes |
| Completeness | Every spec requirement addressed | Missing sections, skipped edge cases |
| Internal consistency | No self-contradiction | Section A says X, section B says not-X |
| Technical correctness | Code compiles/runs, logic sound | Syntax/logic errors, wrong complexity claims |
| Compliance | Project constraints honored | Banned terms, tone/regulatory violations |
Extend per domain (code: type safety, error handling, secrets, input validation; content: brand voice, no trademark misuse; regulated: required disclaimers, no unsubstantiated guarantees).
Both reviewers must PASS. No partial credit. If either FAILs, merge and dedupe both reviewers'
critical_issues, fix those (suggestions are optional), and re-review.
| Mode | Symptom | Mitigation |
|---|---|---|
| Infinite loop | New issues every round | 3-round cap → escalate |
| Rubber-stamping | Both pass everything | Adversarial prompt: "find problems, default to FAIL" |
| Subjective drift | Flags style, not defects | Tight, objective rubric only |
| Fix regression | Fix A introduces B | Fresh reviewers each round catch it |
| Shared blind spot | Both miss the same thing | Use different models/lenses; add a 3rd reviewer or human spot-check for critical output |
~2–3× generation token cost per cycle. For high-stakes output that's a bargain. For large batches, review a stratified sample (10–15%, min 5), classify failures, fix the whole batch by pattern, then re-sample until a clean sample passes.
verification-loop — deterministic checks (build/lint/test). Run it FIRST; this skill is semantic.council — divergent decision-making among advisors. This skill is convergent pass/fail verification.code-reviewer / security-reviewer agents — the natural reviewer subagents for code output.