YAO Skill 生成审计报告YAO Skill Generation Audit
yao-meta-skill
技能审计报告Yao Meta Skill Audit Report
这份报告默认使用中文简体，把新 Skill 的定位、指标、原理、契约、质量、风险、资产和迭代路线整理为一份可审计的 HTML 报告。Create, refactor, evaluate, and package agent skills from workflows, prompts, transcripts, docs, or notes. Use for skill creation, reusable workflow packaging, skill improvement, evals, and team-ready distribution.
openaiopenaiclaudeclaudegenericgenericvscodevscode
技能概述Overview
先用一屏说明这个 Skill 是什么、给谁用、交付什么。A first-screen explanation of what this skill is, who it serves, and what it delivers.
作用定位Role
- 原始说明可切换到英文查看；默认中文报告保留结论与结构说明。Create, refactor, evaluate, and package agent skills from workflows, prompts, transcripts, docs, or notes. Use for skill creation, reusable workflow packaging, skill improvement, evals, and team-ready distribution.
- 把一次性经验沉淀为可复用、可评估、可迁移的 Skill 包体。Turn one-off experience into a reusable, evaluable, and portable skill package.
- 交付结果：SKILL.md, agents/interface.yaml, skill-ir/examples/yao-meta-skill.json, reports/compiled_targets.md, reports/output_quality_scorecard.md, reports/output_execution_runs.md, reports/output_blind_review_pack.md, reports/output_review_kit.md, reports/output_review_adjudication.md, reports/benchmark_reproducibility.md, reports/output_blind_answer_key.json, reports/conformance_matrix.md, reports/security_trust_report.md, reports/runtime_permission_probes.md, reports/skill_atlas.html, reports/registry_audit.md, reports/package_verification.md, reports/install_simulation.md, reports/upgrade_check.md, reports/adoption_drift_report.md, reports/review_waivers.md, reports/world_class_evidence_plan.md, reports/world_class_evidence_ledger.md, reports/review_annotations.md, reports/review-studio.html, reports/skill-interpretation.html, reports/skill-overview.htmlDeliverables: SKILL.md, agents/interface.yaml, skill-ir/examples/yao-meta-skill.json, reports/compiled_targets.md, reports/output_quality_scorecard.md, reports/output_execution_runs.md, reports/output_blind_review_pack.md, reports/output_review_kit.md, reports/output_review_adjudication.md, reports/benchmark_reproducibility.md, reports/output_blind_answer_key.json, reports/conformance_matrix.md, reports/security_trust_report.md, reports/runtime_permission_probes.md, reports/skill_atlas.html, reports/registry_audit.md, reports/package_verification.md, reports/install_simulation.md, reports/upgrade_check.md, reports/adoption_drift_report.md, reports/review_waivers.md, reports/world_class_evidence_plan.md, reports/world_class_evidence_ledger.md, reports/review_annotations.md, reports/review-studio.html, reports/skill-interpretation.html, reports/skill-overview.html
总览指标Metrics
分数来自本地文件和 reports 证据，缺失时明确标为证据不足。Scores are derived from package files and reports; missing inputs are shown as evidence gaps.
指标判读Reading
先看雷达图判断能力短板，再看每项分数的证据原因。分数不是装饰数字，必须和本地文件、reports 证据或证据不足提示对应。Read the radar first for weak spots, then inspect each score with its evidence. Scores must map to local files, reports, or explicit evidence gaps.
成熟度条Maturity Bar
这里把每个指标压缩为状态、分数和第一条证据，先给出整体判断，再进入下方明细。This compresses each metric into status, score, and the first evidence item before the detailed cards below.
- 稳定Stable完整度Completeness100SKILL.md 已存在，是 Skill 的入口。SKILL.md exists and acts as the skill entrypoint.
- 稳定Stable触发清晰Trigger clarity100frontmatter description 已存在，具备基础路由面。The frontmatter description exists, giving the skill a basic routing surface.
- 稳定Stable证据充分Evidence depth100已生成 20 / 20 类报告证据。Generated 20 / 20 evidence report types.
- 稳定Stable可维护性Maintainability100SKILL.md 约 339 个词/字。SKILL.md is about 339 words/characters.
- 稳定Stable可迁移性Portability100agents/interface.yaml 已存在。agents/interface.yaml exists.
- 关注Watch上下文成本Context cost42入口约 339 个词/字，references 约 16934 个词/字。Entrypoint is about 339 words/characters; references are about 16934.
指标明细Metric Details
完整度Completeness100
- SKILL.md 已存在，是 Skill 的入口。SKILL.md exists and acts as the skill entrypoint.
- README.md 已存在，便于人工阅读。README.md exists for human-readable usage.
- agents/interface.yaml 已存在，便于跨平台适配。agents/interface.yaml exists for cross-platform adaptation.
触发清晰Trigger clarity100
- frontmatter description 已存在，具备基础路由面。The frontmatter description exists, giving the skill a basic routing surface.
- description 有足够长度说明任务边界。The description is long enough to explain the task boundary.
- description 已包含使用场景或排除边界信号。The description includes usage-scenario or exclusion-boundary signals.
证据充分Evidence depth100
- 已生成 20 / 20 类报告证据。Generated 20 / 20 evidence report types.
- skill-ir.json 已存在。skill-ir.json exists.
- compiled_targets.json 已存在。compiled_targets.json exists.
可维护性Maintainability100
- SKILL.md 约 339 个词/字。SKILL.md is about 339 words/characters.
- 入口文件保持克制，可维护性较好。The entrypoint stays restrained, which supports maintainability.
- references/ 已承载扩展指导。references/ carries extended guidance.
可迁移性Portability100
- agents/interface.yaml 已存在。agents/interface.yaml exists.
- manifest.json 已存在。manifest.json exists.
- 目标平台或 adapter target 已声明。Target platforms or adapter targets are declared.
上下文成本Context cost42
- 入口约 339 个词/字，references 约 16934 个词/字。Entrypoint is about 339 words/characters; references are about 16934.
- 分数越高代表上下文成本越低。A higher score means lower context cost.
- 上下文成本偏高，建议压缩入口或拆分 references。Context cost is high; compress the entrypoint or split references further.
能力画像Capability
判断这个 Skill 在能力地图中的位置和复用范围。Places this skill on a capability map and clarifies reuse scope.
画像摘要Profile
- 能力类型：Execution operationCapability type: Execution operation
- 成熟度：governedMaturity: governed
- 触发强度：手动触发 + description 路由Trigger strength: Manual activation plus description-based routing
- 复用范围：跨平台Reuse scope: Cross-platform
原理结构Principle
说明入口、参考、脚本、评估和报告如何组成一个稳定闭环。Explains how entrypoint, references, scripts, evals, and reports form a stable loop.
执行流程Execution Flow
- 原始说明可切换到英文查看；默认中文报告保留结论与结构说明。For one-off/no reusable process: `Do not create a skill`; `near-neighbor`; require `repeated use` + `reusable output contract`.
- 原始说明可切换到英文查看；默认中文报告保留结论与结构说明。Capture job, output, exclusions, constraints, standards, and the lightest fit.
- 原始说明可切换到英文查看；默认中文报告保留结论与结构说明。Scan references in order: external benchmark, user source, local fit; surface only uncertainty or conflict.
- 原始说明可切换到英文查看；默认中文报告保留结论与结构说明。Write `description` early, test route quality, then add only earned folders and gates.
- 原始说明可切换到英文查看；默认中文报告保留结论与结构说明。Add output-risk, artifact-design, prompt-quality, system-model, and next directions only when useful.
调用方式How To Use
- 当你需要把工作流或笔记沉淀成结构精简、触发清晰且带必要 eval 的可复用 Skill 时使用 $yao-meta-skill。Use $yao-meta-skill to turn my workflow or notes into a reusable skill with lean structure, clear triggering, and the right evals.
- 当用户请求与该 Skill 的触发描述匹配时使用。Use this skill when the request matches the frontmatter description.
契约边界Contract
把触发、输入、输出和排除场景放在同一屏。Keeps trigger, inputs, outputs, and exclusions on the same screen.
触发描述Trigger
该 Skill 的触发描述来自 SKILL.md frontmatter；默认中文报告先呈现能力边界，原始英文描述可切换到英文查看。Create, refactor, evaluate, and package agent skills from workflows, prompts, transcripts, docs, or notes. Use for skill creation, reusable workflow packaging, skill improvement, evals, and team-ready distribution.
输入材料Inputs
- 原始说明可切换到英文查看；默认中文报告保留结论与结构说明。rough workflow notes, SOPs, runbooks, prompts, transcripts, documents, or repeated task descriptions
- 原始说明可切换到英文查看；默认中文报告保留结论与结构说明。an existing skill directory that needs refactor, evaluation, packaging, or governance hardening
- 原始说明可切换到英文查看；默认中文报告保留结论与结构说明。target platform requirements such as OpenAI, Claude, generic Agent Skills, or team distribution
- 原始说明可切换到英文查看；默认中文报告保留结论与结构说明。benchmark references, local constraints, desired maturity tier, and review standards
输出结果Outputs
- 原始说明可切换到英文查看；默认中文报告保留结论与结构说明。A working skill package with lean SKILL.md, aligned agents/interface.yaml, justified references, scripts only when useful, eval evidence, reports, packaging metadata, and clear next iteration recommendations.
- 结构化 Skill 目录，共 8 类关键资产。Structured skill directory with 8 key asset groups.
不应触发Should Not Trigger
- 只需要一次性回答、没有复用价值的临时请求。One-off requests that do not need reusable skill behavior.
- 要求直接执行相邻任务，而不是沉淀或使用这个 Skill。Requests to perform an adjacent task directly rather than create or use this skill.
- 缺少必要事实且用户不允许澄清的场景。Cases that lack required facts and do not allow clarification.
质量评估Quality
展示强项、缺口和建议，避免只给分不解释。Shows strengths, gaps, and recommendations instead of scores without explanation.
| 类型Type | 证据Evidence | 建议Action |
|---|---|---|
| 强项Strength | 触发面保持精简，并锚定在 frontmatter description。The trigger surface stays lean and anchored in the frontmatter description. | 保留并复用Keep |
| 强项Strength | 已生成 Skill IR，核心语义可先于平台打包被审查和迁移。Skill IR is generated so core semantics can be reviewed and migrated before platform packaging. | 保留并复用Keep |
| 强项Strength | 已生成目标编译报告，可审查 IR 到 OpenAI、Claude、generic 等目标契约的映射。Target compilation evidence is generated to review how IR maps to OpenAI, Claude, generic, and other target contracts. | 保留并复用Keep |
| 缺口Gap | 上下文成本需要补强：入口约 339 个词/字，references 约 16934 个词/字。Context cost needs improvement: Entrypoint is about 339 words/characters; references are about 16934. | 纳入下一轮修复Fix next |
执行证据Execution Evidence
- 变体运行：10Variant runs: 10
- 模型执行：10Model executions: 10
- 记录样本：0Recorded fixtures: 0
- Token 估算：0Token estimates: 0
盲评审定Blind Adjudication
- 评审进度：0 / 5Review progress: 0 / 5
- 待评审：5Pending review: 5
- 一致率：暂无Agreement rate: not available yet
- 非法决策：0Invalid decisions: 0
评审原则Review Rule
- 先记录 reviewer 对 A/B 的选择，再打开答案 key 计算一致率。Record the reviewer's A/B choice before opening the answer key and calculating agreement.
- 缺少真实 reviewer 决策时只显示待评审，不伪造人工结论。When real reviewer decisions are missing, show pending status instead of fabricating human conclusions.
运行原则Run Rule
- recorded fixture 只能证明可复现样本，不等同于模型执行。A recorded fixture proves reproducible samples only; it is not model execution.
- 只有 provider runner 返回 model metadata 时才计入 model-executed。Only provider runners that return model metadata count as model-executed.
风险治理Risk
提前暴露误触发、漂移、证据不足和迁移风险。Surfaces trigger, drift, evidence, and portability risks before the package grows.
| 风险Risk | 信号Signal | 应对Response |
|---|---|---|
| 误触发风险Trigger risk | frontmatter description 已存在，具备基础路由面。The frontmatter description exists, giving the skill a basic routing surface. | 先补证据和边界，再增加包体复杂度。Improve evidence and boundaries before adding package complexity. |
| 输出漂移风险Output drift risk | 已生成 20 / 20 类报告证据。Generated 20 / 20 evidence report types. | 先补证据和边界，再增加包体复杂度。Improve evidence and boundaries before adding package complexity. |
| 证据不足风险Evidence gap risk | 已生成 20 / 20 类报告证据。Generated 20 / 20 evidence report types. | 先补证据和边界，再增加包体复杂度。Improve evidence and boundaries before adding package complexity. |
| 包体膨胀风险Package bloat risk | SKILL.md 约 339 个词/字。SKILL.md is about 339 words/characters. | 先补证据和边界，再增加包体复杂度。Improve evidence and boundaries before adding package complexity. |
| 跨平台迁移风险Portability risk | agents/interface.yaml 已存在。agents/interface.yaml exists. | 先补证据和边界，再增加包体复杂度。Improve evidence and boundaries before adding package complexity. |
世界证据World Evidence
世界级证据尚未完成：4 项待补，0 项已接受。World-class evidence is not complete: 4 pending, 0 accepted.
外部证据External evidence
提供商留出Provider Holdout
缺少真实 provider 模型运行和 token metadata。Missing a real provider model run and token metadata.
阻塞检查Blocked Checks
当前没有阻塞检查。No blocked checks right now.
人工证据Human evidence
人工盲评Human Adjudication
盲评 pair 仍待真实 reviewer 决策。Blind-review pairs still need real reviewer decisions.
阻塞检查Blocked Checks
- 无待判定No pending decisions
- 盲评完成Judgments complete
- Reviewer metadataReviewer metadata
外部证据External evidence
原生权限Native Permission
原生 runtime enforcement 仍待目标客户端或外部安装器证明。Native runtime enforcement still needs target-client or external-installer proof.
阻塞检查Blocked Checks
- 原生执行Native enforcement
外部证据External evidence
原生遥测Native Telemetry
真实外部客户端 metadata-only 事件仍未导入。Real external-client metadata-only events have not been imported yet.
阻塞检查Blocked Checks
- 外部事件External events
包体资产Assets
让 reviewer 快速确认关键文件、目录和资产分布。Lets reviewers confirm key files, directories, and asset distribution quickly.
| 路径Path | 作用Role | 类型Type |
|---|---|---|
| SKILL.md | Skill 入口文件Skill entrypoint | 文件file |
| README.md | 人类可读使用说明Human-readable usage guide | 文件file |
| agents/interface.yaml | 跨平台接口元数据Neutral interface metadata | 文件file |
| manifest.json | 生命周期与打包元数据Lifecycle and portability metadata | 文件file |
| references | 扩展指导与复用资料Extended guidance and reusable notes | 目录folder |
| scripts | 确定性脚本或本地工具Deterministic helpers or local tooling | 目录folder |
| evals | 触发与质量检查Trigger and quality checks | 目录folder |
| reports | 生成的证据与总结报告Generated evidence and overview artifacts | 目录folder |
迭代路线Roadmap
把下一步升级收束为少数高价值动作。Keeps next iteration moves focused and actionable.
补齐世界证据Close world-class evidence
世界级证据仍有 4 项待补；公开完成态 claim 必须继续保持阻塞。World-class evidence still has 4 pending item(s); public completion claims must stay blocked.
- 补齐提供商留出证据：缺少真实 provider 模型运行和 token metadata。Close provider holdout evidence: Missing a real provider model run and token metadata.
- 补齐人工盲评证据：盲评 pair 仍待真实 reviewer 决策。Close human adjudication evidence: Blind-review pairs still need real reviewer decisions.
- 按 ledger 总量继续补齐 4 项待补证据（外部 3 项、人工 1 项；未展开 2 项），并保持 claim guard 为 pending 状态。Close all 4 pending evidence item(s) in the ledger (external 3, human 1; 2 not shown here) and keep the claim guard pending.
全部外部/人工证据被 ledger 接受后，才能进入公开 world-class claim 复核。Only after the ledger accepts all external and human evidence should the public world-class claim move to review.
有选择地借鉴一个成熟模式Borrow one proven pattern on purpose
已经有公开 benchmark 对象，下一步应主动选择一个模式借鉴，而不是松散吸收所有做法。You already have public benchmark objects. The next gain is to choose one pattern intentionally instead of absorbing everything loosely.
- 阅读 `obra/superpowers` 中最值得借鉴的模式。Read the strongest pattern from obra/superpowers.
- 先判断要借鉴的是方法、结构、执行方式还是可迁移性，并且第一轮只借鉴其中一个。Decide whether to borrow method, structure, execution, or portability, but only one of them first.
- 记录本轮不借鉴的内容，避免包体过重。Record what you will not borrow so the package stays light.
包体形态更清晰，也减少偶然过度设计。A cleaner package shape with less accidental over-design.
加固跨环境语义Harden portability semantics
这个 Skill 已经面向跨环境复用，因此早期就需要把契约语义说清楚。The skill already signals reuse across environments, so contract clarity matters early.
- 确认激活模式、执行上下文和信任假设。Confirm activation mode, execution context, and trust assumptions.
- 补充或复核非原生目标端的降级策略。Add or review degradation strategy for non-native targets.
- 至少打包一次 Skill，用来验证 adapter 预期。Package the skill once to verify adapter expectations.
跨环境复用更安全，目标漂移更少。Safer cross-environment reuse with less target drift.