Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Multi-agent adversarial-verification orchestration for high-confidence conclusions. Fan-out finders, then verify every finding through a three-prism panel (exploitability / correctness / refutation) that defaults to disbelief, gate fixes behind load-bearing proof tests that catch agents who falsely claim "done/fixed", and roll out behind a build-first exit-code guard. Use when a fan-out task must produce trustworthy results — security audit, code review, research synthesis, migration — and a sin
.claude/skills/telagod-orchestrating-adversarial-reviews/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-01 | ✗→✓ | ▲ Improved | 41% | 0% |
| case-08 | ✗→✓ | ▲ Improved | 42% | 0% |
| case-10 | ✗→✓ | ▲ Improved | 24% | 0% |
| case-15 | ✗→✓ | ▲ Improved | 45% | 0% |
| case-16 | ✗→✓ | ▲ Improved | 48% | 0% |
<!-- safety-scan: ignore RM_RF_ROOT,CURL_PIPE_SH,PROMPT_INJECTION 本 skill 把危险命令(| tail 吞退出码、docker rm 误删、agent 谎报)列为反模式教学，自身不执行 --> <!-- safety-scan: ignore TOOLS_PRIVILEGED 知识型 skill，仅 Read；文中 Workflow / docker / go / git 命令由 agent 自有工具执行，非本 skill 落盘运行 -->
> 单个 agent 会谎报"已修复 / 全覆盖 / 没问题"。结论的可信度不来自"谁说的"，来自"扛过几次推翻"。 > 本 skill 是编排骨架：fan-out 发现 → 三棱镜对抗验证 → 证明性 guard → 守卫式上线。 > 信级：运行时行为 / 证明测试 > 多 agent 多数裁决 > 单 agent 自报（永远 [unverified]）。
| 场景 | 用 | 理由 | |------|----|------| | 授权安全审计 / 加固闭环 | ✅ | 首个范例，见 workflow | | 大面积代码审查（多维度、需高可信） | ✅ | dimensions → find → 对抗验证 | | 研究综合 / 事实核查（结论要扛得住） | ✅ | 多源 fan-out + 证伪棱镜 | | 大规模迁移 / 重构（site 发现 + 逐项验证） | ✅ | pipeline 逐项独立 + 证明测试 |
shipping-changes 的"何时不使用"）。securing-systems / analyzing-security，本 skill 不重写知识，只编排。Recon (fan-out) 每维一个 finder, 并行深读, schema 出结构化 findings
| pipeline 而非 barrier: 维度A的发现可在维度B还在找时就进验证
v
Verify (三棱镜) 每条 finding 派 N 个 verifier, 各执一镜, 默认怀疑
| 可利用性 / 正确性 / 证伪猎杀 —— 票数 >= 多数 才保留
v
Synthesize / Ship 合成定级报告; 若是修复任务 -> 证明测试 guard -> build-first 上线对应 Workflow 工具的 pipeline(items, findStage, verifyStage)（默认无栅栏，墙钟最短）。仅当"下一阶段需全部上一阶段结果"（去重 / 早退 / 跨条比较）才用 parallel 栅栏。
confirmed 才保留。cmd | tail 会吞掉 cmd 的退出码，判码用 cmd > log 2>&1; rc=$?。> 每条护栏的细节、PoC 判据、反向证伪实操、安全审计 worked example，见 references/workflow.md。
up --build 2>&1 | tail && rm <old>——管道吞码，build 失败仍删旧件。isolation: worktree。securing-systems —— 攻防知识总路由（找什么洞）。shipping-changes —— 单上下文变更闭环脊柱。cultivating-skills —— 本 skill 的孵化器 / 安全脊柱 / 升级漏斗。| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-21 | pass→pass | 12,077 | 9,969 | -17% | 1 | 1 | 0% | 1,615 | 2,797 | +73% | 0 | 0 | — |
case-03 | fail→fail | 15,666 | 18,186 | +16% | 1 | 1 | 0% | 2,646 | 3,724 | +41% | 0 | 0 | — |
case-01 | fail→pass | 18,086 | 24,684 | +36% | 1 | 1 | 0% | 3,199 | 4,520 | +41% | 0 | 0 | — |
case-02 | fail→fail | 36,958 | 4,548 | -88% | 1 | 1 | 0% | 6,163 | 1,871 | -70% | 0 | 0 | — |
case-04 | pass→fail | 7,380 | 7,617 | +3% | 1 | 1 | 0% | 1,060 | 2,455 | +132% | 0 | 0 | — |
case-05 | pass→pass | 18,427 | 18,377 | -0% | 1 | 1 | 0% | 2,953 | 4,390 | +49% | 0 | 0 | — |
case-06 | pass→pass | 5,437 | 5,735 | +5% | 1 | 1 | 0% | 710 | 2,181 | +207% | 0 | 0 | — |
case-07 | pass→pass | 14,573 | 12,218 | -16% | 1 | 1 | 0% | 2,347 | 3,251 | +39% | 0 | 0 | — |
case-08 | fail→pass | 14,482 | 14,177 | -2% | 1 | 1 | 0% | 2,613 | 3,708 | +42% | 0 | 0 | — |
case-09 | pass→pass | 17,175 | 12,889 | -25% | 1 | 1 | 0% | 2,679 | 3,127 | +17% | 0 | 0 | — |
case-10 | fail→pass | 10,064 | 4,161 | -59% | 1 | 1 | 0% | 1,506 | 1,865 | +24% | 0 | 0 | — |
case-11 | pass→pass | 7,880 | 4,728 | -40% | 1 | 1 | 0% | 1,220 | 1,956 | +60% | 0 | 0 | — |
case-12 | pass→pass | 11,854 | 6,913 | -42% | 1 | 1 | 0% | 1,858 | 2,338 | +26% | 0 | 0 | — |
case-13 | pass→pass | 15,173 | 10,306 | -32% | 1 | 1 | 0% | 2,177 | 2,820 | +30% | 0 | 0 | — |
case-14 | pass→pass | 15,878 | 12,289 | -23% | 1 | 1 | 0% | 2,595 | 3,238 | +25% | 0 | 0 | — |
case-15 | fail→pass | 14,451 | 12,966 | -10% | 1 | 1 | 0% | 2,159 | 3,140 | +45% | 0 | 0 | — |
case-16 | fail→pass | 13,670 | 12,629 | -8% | 1 | 1 | 0% | 2,203 | 3,251 | +48% | 0 | 0 | — |
case-17 | pass→pass | 12,629 | 7,549 | -40% | 1 | 1 | 0% | 1,979 | 2,390 | +21% | 0 | 0 | — |
case-18 | fail→pass | 10,562 | 6,092 | -42% | 1 | 1 | 0% | 1,574 | 2,250 | +43% | 0 | 0 | — |
case-19 | fail→pass | 26,155 | 11,737 | -55% | 1 | 1 | 0% | 1,104 | 3,154 | +186% | 0 | 0 | — |
case-20 | pass→pass | 17,370 | 16,491 | -5% | 1 | 1 | 0% | 2,723 | 3,995 | +47% | 0 | 0 | — |
case-22 | fail→pass | 9,737 | 7,193 | -26% | 1 | 1 | 0% | 1,722 | 2,533 | +47% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted, and 21 counted toward the lift figure. The other 1 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +32 percentage points is the difference between those two pass rates over the 21 comparable cases. 2 cases got worse with the skill loaded, and they are included in that figure.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.