Install any skill in seconds. Free to start, no credit card required.
Get Started Free →能力蒸馏工作流——从用户批准的强模型访谈或真实任务轨迹中提取非显然的判断规则,形成可审计的 judgment packet,再交给 skill-audit 和 skill-creator 决定是否落成可加载 skill。当用户说“蒸馏这个模型的判断力”“把这次任务的关键决策固化下来”“模型窗口要关了,保留它在某类场景的判断”时使用。普通流程文档、直接写 SKILL.md、泛化最佳实践或未授权的会话日志扫描不使用本 skill。
.claude/skills/majiayu000-capability-distill/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-05 | ✗→✓ | ▲ Improved | 1% | 0% |
| case-16 | ✗→✓ | ▲ Improved | 106% | 0% |
| case-06 | ✗→✓ | ▲ Improved | -15% | 0% |
| case-08 | ✗→✓ | ▲ Improved | 139% | 0% |
| case-09 | ✗→✓ | ▲ Improved | 152% | 0% |
把“为什么在两个合理选项中选择其中一个”提取成证据支持的判断规则。不要把操作步骤、通用建议或原始会话内容换个格式包装成 skill。
capability-distill:选择获准证据、还原决策事件、提取判断规则、输出 judgment packet。skill-audit:判断这些规则是否值得成为 skill、归属哪个现有 skill、是否重复以及如何分层。skill-creator:创建或修改具体的 SKILL.md、设计 with-skill/baseline eval、迭代和验证触发描述。skill-lifeguard:当产物属于高影响工作流时补可靠性契约和漂移修复闭环。不要在本 skill 中复制后三者的完整写作、注册或评测流程。目标是提供它们可消费的高信号输入。
rule_id 回写到下一轮 judgment packet,而不是只改措辞。在读取额外材料前记录以下字段:
yamlsource_scope: approved_roots: [] approved_artifact_types: [] external_model_destination: raw_content_authorized: false output_path:
当前对话和用户本次明确附带的文件可直接使用。其他路径、历史日志和 memory 不因“可能有帮助”而自动进入范围;缺少批准时先询问并暂停对应读取。
执行以下数据纪律:
~/.claude、~/.codex 或任何固定运行时路径存在。使用用户给出的路径、当前工作区和当前运行时可用的搜索工具。先区分对象:
skill-audit / skill-creator,不做判断蒸馏。对候选场景记录 judgment_density、evidence_strength 和 existing_coverage(high / medium / low)。只保留判断密度高、证据至少中等且现有覆盖不完整的场景。
从获准证据中提取事件摘要,而不是复制原文。每个事件至少包含:
yamldecision_event: evidence_ref: context: viable_options: [] chosen_option: observed_signal: outcome: counterfactual: redactions_applied: []
evidence_ref 使用本地、非敏感的定位信息,例如仓库相对路径和 commit SHA;不要把原始聊天文本塞进该字段。没有 outcome 的事件只能作为假设,不能升级为高置信规则。
在写任何 judgment_rule 前,先在 source_scope 获准的 roots 或已附 inventory 内搜索现有 rules、skills、registry 和相近 owner;先查索引和元数据,再读最可能相关的最小片段。没有获准的 owner 根目录或清单时,先请求范围并暂停规则起草,不能把“未提供”当作“无覆盖”。
每个候选场景记录 searched_roots、queries、candidate_owners、逐 owner 的 coverage / evidence_ref、uncovered_gap 和 decision:
covered:停止起草,不生成新规则。partial:只允许为明确的 uncovered_gap 起草增量,默认更新现有 owner。uncovered:才可进入新规则候选;仍须在交付前完成逐规则归属复核。该记录是进入下一步的硬 gate。后面的 skill-audit 复核不能替代这次写作前审计。
窗口蒸馏时,对每个脱敏场景逐个询问强模型:
轨迹蒸馏时,围绕已有事件回答相同问题,并比较成功与失败轨迹。不要让模型“一次写完整指令体系”;那会掩盖证据和规则之间的映射。
每条规则使用以下结构:
yamljudgment_rule: rule_id: scenario: observable_signal: default_action: exception: stop_or_escalate: evidence_refs: [] confidence: open_questions: []
规则不得包含源材料中的秘密值或大段原文。confidence 由证据数量、结果可观察性和反例覆盖决定;单一无结果片段不能标 high。
逐条运行三关,并记录被删除的 rule_id 与原因:
规则通过三关仍需具备 observable_signal、exception 和 stop_or_escalate;缺一项就返回提取阶段,不用“按情况判断”填空。
把通过前置重叠审计和三关过滤的 packet 交给 skill-audit,要求它对每条规则复核:existing_owner、coverage、recommended_home。完整覆盖的规则删除;部分覆盖的规则优先更新原 skill;只有明确无归属的高信号规则才进入新 skill brief。
用户要求可加载 skill 时,再调用 skill-creator:
SKILL.md 所在目录下的 evals/evals.json 作为最低测试集(从 skill 根解析,不从当前仓库根解析),并为目标领域增加真实、脱敏场景。skill-lifeguard 检查负例、checkpoint、done condition、replay hook 和 drift signal。来源强模型或日常目标模型升级时,不直接覆盖旧 packet 或 skill。先复用原先获准的 source_scope、场景清单、decision_event 和 eval;新增证据仍需单独批准。把旧版和候选新版并列保存,并记录可验证的来源标签、模型版本和日期;未知字段留空。
按稳定的 rule_id 做逐条 diff;旧版没有稳定 ID 时,使用 scenario + observable_signal 对齐。每条差异标为 unchanged、refined、added 或 removal_candidate,并附证据、行为影响和未决问题。diff 只是审计产物,不自动证明新版更好,也不授权删除或覆盖。
removal_candidate。skill-audit / skill-creator;经用户确认后更新原 owner,不另建同场景 skill。保留旧版本或 commit 作为 rollback,不自动删除历史产物。packet_only 仅在以下条件全部满足时完成:
source_scope 已记录,所有读取和外部发送均在批准范围内。decision_event,并具有信号、默认动作、例外和停止/上抛条件。open_questions。skill-audit 已给出保留、更新现有 skill 或不创建的归属结论。skill_delivery 还必须满足:
skill-creator 的真实 with-skill/baseline eval 已运行并保存结果。| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | pass→fail | 13,758 | 12,304 | -11% | 1 | 1 | 0% | 2,118 | 4,155 | +96% | 0 | 0 | — |
case-15 | fail→fail | 12,419 | 7,912 | -36% | 1 | 1 | 0% | 1,979 | 3,591 | +81% | 0 | 0 | — |
case-02 | fail→fail | 20,046 | 4,451 | -78% | 1 | 1 | 0% | 2,673 | 2,749 | +3% | 0 | 0 | — |
case-03 | pass→fail | 29,785 | 35,699 | +20% | 1 | 1 | 0% | 4,726 | 2,883 | -39% | 0 | 0 | — |
case-04 | fail→fail | 8,671 | 11,163 | +29% | 1 | 1 | 0% | 1,405 | 3,040 | +116% | 0 | 0 | — |
case-05 | fail→pass | 24,039 | 10,417 | -57% | 1 | 1 | 0% | 4,078 | 4,118 | +1% | 0 | 0 | — |
case-16 | fail→pass | 10,652 | 5,693 | -47% | 1 | 1 | 0% | 1,574 | 3,248 | +106% | 0 | 0 | — |
case-06 | fail→pass | 24,799 | 8,425 | -66% | 1 | 1 | 0% | 4,635 | 3,952 | -15% | 0 | 0 | — |
case-07 | pass→pass | 8,190 | 5,038 | -38% | 1 | 1 | 0% | 1,416 | 3,131 | +121% | 0 | 0 | — |
case-08 | fail→pass | 9,329 | 8,008 | -14% | 1 | 1 | 0% | 1,528 | 3,645 | +139% | 0 | 0 | — |
case-09 | fail→pass | 9,492 | 7,149 | -25% | 1 | 1 | 0% | 1,370 | 3,453 | +152% | 0 | 0 | — |
case-10 | pass→fail | 9,718 | 3,698 | -62% | 1 | 1 | 0% | 1,399 | 2,950 | +111% | 0 | 0 | — |
case-11 | fail→fail | 11,876 | 9,069 | -24% | 1 | 1 | 0% | 1,902 | 3,883 | +104% | 0 | 0 | — |
case-12 | fail→pass | 6,974 | 5,631 | -19% | 1 | 1 | 0% | 1,180 | 3,335 | +183% | 0 | 0 | — |
case-13 | pass→pass | 13,329 | 19,573 | +47% | 1 | 1 | 0% | 2,006 | 5,513 | +175% | 0 | 0 | — |
case-14 | pass→pass | 9,880 | 10,713 | +8% | 1 | 1 | 0% | 1,667 | 4,127 | +148% | 0 | 0 | — |
case-17 | pass→pass | 36,840 | 13,408 | -64% | 1 | 1 | 0% | 2,536 | 4,592 | +81% | 0 | 0 | — |
case-18 | fail→fail | 7,265 | 7,087 | -2% | 1 | 1 | 0% | 1,092 | 3,520 | +222% | 0 | 0 | — |
case-19 | fail→pass | 5,170 | 9,403 | +82% | 1 | 1 | 0% | 767 | 3,983 | +419% | 0 | 0 | — |
case-20 | pass→pass | 5,298 | 6,162 | +16% | 1 | 1 | 0% | 802 | 3,458 | +331% | 0 | 0 | — |
case-21 | pass→fail | 16,031 | 3,706 | -77% | 1 | 1 | 0% | 3,079 | 2,859 | -7% | 0 | 0 | — |
case-22 | pass→pass | 12,999 | 6,357 | -51% | 1 | 1 | 0% | 2,042 | 3,474 | +70% | 0 | 0 | — |
case-23 | pass→pass | 7,950 | 4,960 | -38% | 1 | 1 | 0% | 1,169 | 3,242 | +177% | 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. 23 cases were attempted, and 22 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 +13 percentage points is the difference between those two pass rates over the 22 comparable cases. 4 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.