360 Secure SkillHub
360 安全技能中心

Data Knowledge Architect — Model the Meaning Behind Your Data

该技能把 REST/OpenAPI、CLI 或表/DDL 契约作为唯一事实来源,经四阶段:摄取并分类操作、HTML 评审原子决策、导出 Google OKF 或仓库 YAML、结构/语义校验。适用于企业知识、语义层、维度模型或本体建模。输入接口契约,输出受治理语义层与可审阅报告;坚持证据优先、先定粒度,缺失证据会阻塞并逐问确认。

agenticweb4
agenticweb4
热度 33.02
最近更新 2026-09-30 10:54:44
分类:知识依赖:读写文件 / 内部调用大模型 / 代码执行应用场景:数据处理 / API 集成 / 报告撰写 / 模板

安装技能

复制下面的话术发给你的 Agent,它会通过 secure-skills 检索并安装该技能。

执行命令安装此 skill: npx secure-skills-cli add data-knowledge-architect
Metadata
name
data-knowledge-architect
description
Builds semantic models from API, CLI, and database contracts with an HTML review and OKF/YAML export. Use when modeling enterprise knowledge, semantic layers, dimensional models, or ontology.
metadata
{"language":"en","translation_of":"data-knowledge-architect-cn","version":"1.1.0","openclaw":{"homepage":"https://github.com/ontology-of-everything/concept-skills/tree/main/skills/en/data-knowledge-architect"}}

Data Knowledge Architect — Model the Meaning Behind Your Data

Turn an interface contract—REST/OpenAPI, CLI, or table/DDL—into a governed semantic layer in four strictly sequential phases. Advance only when the current phase meets its exit condition.

Phase 1 Ingest → Phase 2 Review → Phase 3 Emit → Phase 4 Verify
   ingest.md       review.md      emit-okf.md      verify.md
                                  emit-yaml.md

The interface is the sole source of truth. Never invent fields, grain, enums, or values. Stop on missing evidence and ask one question at a time.

Workflow

  1. Ingest (references/ingest.md) — Normalize interfaces into a complete operation inventory and classify each operation as fact or dimension lookup with read/write status. Exit when every operation is classified.
  2. Review (references/review.md) — Decompose the model into atomic decisions. Show objective, options, evidence, risks, dependencies, and tradeoffs, with fixed YAGNI guidance for facts, dimensions, measures, and routing. Inject model JSON into assets/review-template.html; record feedback in amendments.md and regenerate. Stop after the report. Exit only on approved:true with no open or blocked decisions.
  3. Emit — Default to a Google OKF v0.1 bundle using references/emit-okf.md; use repository YAML only when explicitly requested, following references/emit-yaml.md. Both targets share field semantics. Generate a thin root index for two or more bundles. Exit when every artifact exists.
  4. Verify (references/verify.md) — Run structural checks, semantic lint, and OKF hard constraints when applicable. Report every pass and failure. Return non-mechanical failures to Phase 2 with one blocking question.

Critical Rules

  1. Evidence only — Unsupported fields, grain, enums, or values become TODO(verify) and an open decision.
  2. Grain first — Attach no dimensions or measures before grain is fixed.
  3. One blocking ask — Stop only when an answer changes the model; ask one question with options. Use the HTML review for bulk decisions.
  4. Layer split — Shape, routing, and grain belong to the semantic layer. Values typed into commands or read from responses—enums, codes, paths, and commands—belong to the contract layer; the semantic layer only points to them.
  5. Stable names — Do not rename confirmed fact or dimension names; OKF concept IDs derive from paths.
  6. Decision gate — Every decision starts pending. Phase 2 must stop after producing the report; Phase 3 requires approved:true and no open or blocked item.

Load by phase: ingest.md, review.md, emit-okf.md or emit-yaml.md, then verify.md. Read references/examples.md for an end-to-end example.