Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Use when explicit CrossFrame output needs review for reasoning fidelity, evidence boundaries, source anchors, concept drift, article collapse, or repair steps.
.claude/skills/sickn33-crossframe-review/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-02 | ✗→✓ | ▲ Improved | 58% | 0% |
| case-07 | ✗→✓ | ▲ Improved | 154% | 0% |
| case-06 | ✗→✓ | ▲ Improved | 137% | 0% |
| case-10 | ✗→✓ | ▲ Improved | 197% | 0% |
| case-16 | ✗→✓ | ▲ Improved | 86% | 0% |
crossframe-suite routes a CrossFrame output into the review gate.This AAS-ready copy preserves the original CrossFrame skill body below. Chinese remains the canonical semantic layer; English metadata is only for discovery, installation, and repository review.
crossframe-suite routing; do not apply it as a generic default reasoning layer.如果评审对象来自多个 CrossFrame skill 的连续工作流,先读取 ../crossframe-suite/SKILL.md 还原应有调度链,再判断是否有漏触发、误触发或跳过质量闸。
本 skill 只做评审与修复建议,不替代 crossframe 生成诊断,也不替代 crossframe-essay 写文章。中文为权威语义;英文只作 skill id、文件名和接口说明。
每次触发后,先读取 canonical skill,而不是复制它们的正文:
../crossframe/SKILL.md。../crossframe/references/read-routing-map.md。../crossframe/references/runtime-read-policy.md 与 ../crossframe/references/continuity-closure-map.md,判断本应触发哪些 v5.0 连续联读包;只有需要包说明、源锚点或闭包细节时,再定向读取 ../crossframe/references/continuity-bundles.md 或具体包文件。../crossframe/templates/read-state-capsule.md 规定的 v5-read-state-capsule 是否存在并被下游复用。../crossframe/worksheets/source-continuity-check.md 与 ../crossframe/worksheets/source-anchor-integrity-check.md,检查闭包是否完整、中心命题和行动边界是否能回指胶囊源锚点。../crossframe-essay/SKILL.md。protocols/review-protocol.md 和 templates/review-report.md。protocols/article-review-protocol.md。../crossframe/references/source-ledger-workflow.md,检查来源台账字段是否完整。references/ 中的评分表、失败类型表和证据边界清单。不要把 CrossFrame 主 skill、文章 skill、eval、examples 或完整案例复制到本 skill 输出中。评审时只引用必要规则名、触发点和证据位置。若 v5-read-state-capsule 已存在,下游默认复用胶囊,不得为了评审而重复整块读取源索引。
判断一个输出是否真的完成了 CrossFrame 的最低推理链:
以下问题一旦出现,要在评审中明确定位;严重时直接判为不合格:
continuity-bundles.md 的联读包,却只读单个概念卡、单个 protocol 或单个摘录就下判断。crossframe 生成 v5-read-state-capsule 的任务没有胶囊,导致 essay/review 各自重读源索引或发明路由。默认使用 templates/review-report.md。最终评审必须包含:
若用户只要一句话结论,也要保留“是否合格 + 主要失败点 + 下一步修复”的最小结构。
当本 skill 经 crossframe-suite 作为默认质量闸调用,而用户没有明确要求“只要评审/完整评审报告/不要文章”时,评审不接管最终输出:
结构洞察底稿 与 文章正文。# 结构洞察底稿 + # 文章正文,最多追加一行短质量闸摘要。templates/review-report.md 只在用户显式要求完整评审报告、只评审已有输出,或硬失败需要说明阻断原因时作为主输出。评分只是辅助,硬失败优先:
触发人格审判、伪造引用、跳过文章底稿、强判断越级、证据边界完全缺失、连续性保真失败时,即使文字流畅,也不能判为合格。
评审输出优先给可执行修复,不默认重写全文。除非用户要求“直接改写”,否则只给:
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | pass→fail | 16,487 | 15,555 | -6% | 1 | 1 | 0% | 2,934 | 4,798 | +64% | 0 | 0 | — |
case-02 | fail→pass | 19,764 | 16,902 | -14% | 1 | 1 | 0% | 3,613 | 5,717 | +58% | 0 | 0 | — |
case-07 | fail→pass | 14,440 | 6,709 | -54% | 1 | 1 | 0% | 1,401 | 3,557 | +154% | 0 | 0 | — |
case-03 | pass→fail | 8,070 | 6,671 | -17% | 1 | 1 | 0% | 1,304 | 3,460 | +165% | 0 | 0 | — |
case-04 | fail→fail | 9,779 | 7,882 | -19% | 1 | 1 | 0% | 1,657 | 3,566 | +115% | 0 | 0 | — |
case-05 | fail→fail | 8,496 | 13,056 | +54% | 1 | 1 | 0% | 1,491 | 4,386 | +194% | 0 | 0 | — |
case-06 | fail→pass | 12,843 | 15,406 | +20% | 1 | 1 | 0% | 2,020 | 4,779 | +137% | 0 | 0 | — |
case-08 | pass→pass | 9,996 | 6,931 | -31% | 1 | 1 | 0% | 1,646 | 3,431 | +108% | 0 | 0 | — |
case-09 | pass→pass | 5,600 | 4,897 | -13% | 1 | 1 | 0% | 938 | 3,090 | +229% | 0 | 0 | — |
case-10 | fail→pass | 7,170 | 5,539 | -23% | 1 | 1 | 0% | 1,103 | 3,275 | +197% | 0 | 0 | — |
case-11 | pass→pass | 9,715 | 8,571 | -12% | 1 | 1 | 0% | 1,493 | 3,725 | +149% | 0 | 0 | — |
case-12 | pass→fail | 9,661 | 4,013 | -58% | 1 | 1 | 0% | 1,705 | 3,103 | +82% | 0 | 0 | — |
case-13 | pass→pass | 11,876 | 9,469 | -20% | 1 | 1 | 0% | 2,172 | 3,905 | +80% | 0 | 0 | — |
case-14 | pass→pass | 6,818 | 4,353 | -36% | 1 | 1 | 0% | 1,254 | 3,046 | +143% | 0 | 0 | — |
case-15 | pass→pass | 7,823 | 4,072 | -48% | 1 | 1 | 0% | 1,396 | 3,119 | +123% | 0 | 0 | — |
case-16 | fail→pass | 11,081 | 7,444 | -33% | 1 | 1 | 0% | 1,938 | 3,612 | +86% | 0 | 0 | — |
case-17 | pass→pass | 9,574 | 8,961 | -6% | 1 | 1 | 0% | 1,673 | 3,765 | +125% | 0 | 0 | — |
case-18 | fail→pass | 7,027 | 4,464 | -36% | 1 | 1 | 0% | 1,245 | 2,976 | +139% | 0 | 0 | — |
case-19 | fail→pass | 8,936 | 5,675 | -36% | 1 | 1 | 0% | 1,330 | 3,344 | +151% | 0 | 0 | — |
case-20 | fail→pass | 10,361 | 4,968 | -52% | 1 | 1 | 0% | 1,785 | 3,203 | +79% | 0 | 0 | — |
case-21 | fail→pass | 6,401 | 5,248 | -18% | 1 | 1 | 0% | 1,049 | 3,273 | +212% | 0 | 0 | — |
case-22 | pass→pass | 5,872 | 1,833 | -69% | 1 | 1 | 0% | 900 | 2,584 | +187% | 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. The headline lift of +27 percentage points is the difference between those two pass rates over the 22 comparable cases. 3 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.