CAPABILITIES × ORCHESTRATION 15 SKILLS · 04 FLOWS

Skills
& Flows

Skill 是能力单元,Workflow 决定它们如何协作。目录与流程都直接来自仓库里的 Markdown。

01 / DIRECTORY

Skill 导航

按仓库中的真实目录分组;展开文件夹查看 README 维护的简短说明。顶层 Skill 保持在 skills/,外部下载的 community/ 完全不发布。

顶层 Skills skills/ 03
  • codex-external-agent-testskill

    仅在明确需要外部进程、登录或跨 runtime 行为时,由 Codex 通过本机缓存配置启动 Claude Code 或 OpenCode 测试一个 Skill;不再配置或启动 Codex target。

  • codex-subagent-testskill

    单 Skill 测试的默认入口;用当前 Codex 会话的原生 subagents 做新上下文 smoke、回归、负向、安全、重复稳定性和有界并发行为测试,不冒充进程或认证级隔离。

  • codex-subagent-workflow

    在当前 Codex 会话内用原生 subagents 编排生产任务;优先采用可信项目 .codex/config.toml 的标准 [agents] 角色,未配置角色时才从任务描述自动拆分,并由主 agent 集成验收。

gamemaker/ skills/gamemaker/ 05
  • build-unity-scene

    读取 Unity 项目架构与关卡需求,按既有边界创建或修改场景并完成验证。

  • configure-unity-mcp

    安装、修复、迁移并完整验证 Codex 与 Unity Editor 的 MCP 集成。

  • discuss-game-design

    基于游戏设计文档讨论具体设计决策,并区分事实、综合、提案与开放问题。

  • play-unity-game

    实际游玩并评估 Unity 游戏或场景,验证玩法循环和复现交互问题。

  • search-game-art

    搜索游戏美术资源并给出经过来源与许可证核验的候选清单,不自动下载或导入。

kgdistiller/ skills/kgdistiller/ 05
  • extract-and-export-notes

    从跨领域 Git 改动提取 Markdown、Typst、LaTeX 候选知识,委托查询和入库后发布网页。

  • extract-paper-concepts

    从网页、DOI 或标题找到规范 PDF,逐页核验并预处理为无图片、图表可追溯的 TeX,再查询个人知识库并生成默认不合并的联邦概念图。

  • ingest-kgdistiller

    调用 kgdistiller 随附的写入 Skill,通过 plan/apply 事务、崩溃恢复和 canonical receipt 把已审查知识写入个人图谱;qlblog 只维护发现入口。

  • query-kgdistiller

    调用 kgdistiller 随附的只读 Skill,批量查询、消歧、GraphRAG 对齐并返回小型证据包;qlblog 只维护发现入口。

  • trace-concept-lineage

    把论文概念批量整理为来源可追溯的概念档案、知识图谱和前置阅读路线。

notes/ skills/notes/ 02
  • create-latex-math-notes

    新建使用 ElegantBook 语法和 LaTeX-to-Typst 适配器的轻量数学笔记项目。

  • create-math-notes

    新建可直接由 VS Code/Tinymist 编辑的 Typst-first 数学课程或专题。

02 / WORKFLOWS.MD

Skill Workflows

单个 Skill 只描述一种可复用能力;这里记录多个 Skill 如何编排成可以交付结果的工具流。流程图和图下文字都是普通 Markdown,可以按需要增删节点、分支和说明,不受页面字段约束。

01
WORKFLOW / MARKDOWN AUTHORED

笔记进入知识库并发布 Web

flowchart LR
    Src["Git 改动<br/>原生 marker"] --> Ext["提取候选图"]
    Ext --> Ask["查询知识库"]
    Ask --> Gate{"身份状态"}
    Gate -->|known| Ref["复用 ref"]
    Gate -->|new / partial| Add["补充缺失知识"]
    Gate -->|uncertain / conflict| Stop["人工审查"]
    Ref --> Put["事务入库"]
    Add --> Put
    Put --> Web["发布 Web"]

extract-and-export-notes 只从 Git 改动和作者写下的 marker 构造候选图;query-kgdistiller 负责识别已有知识,已知项只写 ref;审查通过后由 ingest-kgdistiller 合入个人知识库,成功回执是 Web 发布的前置条件。

遇到 uncertainconflict 时,流程停在审查门前,不猜测身份,也不为了自动发布而创建重复词条。

02
WORKFLOW / MARKDOWN AUTHORED

Codex 原生 Subagent 生产工作流

flowchart LR
    Req["用户任务"] --> Contract["产物 / 权限 / 完成条件"]
    Contract --> Pick["选择最小 Agent Skill 集"]
    Pick --> Config{"可信 .codex/config.toml<br/>存在 agents roles?"}
    Config -->|是| Roles["选择标准配置角色<br/>读取 config_file"]
    Config -->|否| Shape{"按描述生成任务拓扑"}
    Roles --> Lead["Codex Coordinator"]
    Shape -->|单一边界| One["Fresh Codex Subagent"]
    Shape -->|独立或分阶段| Lead
    Lead --> A["Worker A<br/>专属 Skill / 写入边界"]
    Lead --> B["Worker B<br/>专属 Skill / 写入边界"]
    One --> Check["主 Agent 集成与独立验收"]
    A --> Lead
    B --> Lead
    Lead --> Check

codex-subagent-workflow 只使用当前 Codex 会话的 原生 subagents,不通过 shell 或 API 再启动 Codex、Claude Code 或 OpenCode。 主 agent 先固定交付物、权限和完成条件,再检查可信主项目的 .codex/config.toml: 若 [agents.<role>] 已声明角色,就按标准描述、config_file、默认模型/推理强度和并发 限制选择最小角色集;只有没有自定义角色时才根据任务描述自动编排。配置损坏或当前 surface 无法选择已配置角色时明确失败,不悄悄退化。每个 worker 仍需指定最小 Skill、 上下文、写入所有权和依赖关系;独立任务可以有界并发,依赖任务顺序交接。Subagents 共享当前 runtime 与工作区,因此它们是新工作上下文而不是独立安全主体;最终 diff、 产物和 validators 始终由主 agent 集成并验收。这个流程只处理真实生产交付物。

03
WORKFLOW / MARKDOWN AUTHORED

单 Skill 原子与压力测试

flowchart LR
    Skill["一个目标 Skill"] --> Contract["固定 fixture / prompt / invariants"]
    Contract --> Runtime{"明确要求外部进程 / 登录 / runtime?"}
    Runtime -->|否,默认| Copies["新 fixture + 新上下文"]
    Runtime -->|是| External["Claude Code / OpenCode<br/>隔离进程 trial"]
    Copies --> T1["Native Codex Subagent 1"]
    Copies --> T2["Native Codex Subagent N"]
    T1 --> Evidence["逐 trial 产物 / diff / 行为证据"]
    T2 --> Evidence
    External --> Evidence
    Evidence --> Verdict["主 Agent 独立判定"]

codex-subagent-testskill 每个 contract 只测试一个 Skill,是单 Skill 测试默认入口,不承接生产交付。每个 trial 使用新的 fixture copy 和不继承对话历史的原生 Codex subagent;单次 smoke、回归和负向测试使用一个 evaluator,稳定性或压力测试 在可用 subagent slots 内分批运行相同 contract。主 agent 独立检查产物、workspace diff 和 validators,并区分 harness、behavior、artifact、safety 与 orchestration 失败。它提供行为与上下文隔离,不声称全新进程、登录、provider 或文件系统隔离。

只有用户明确要求 Claude Code、OpenCode、跨 runtime 对比或全新进程/认证时,才改用 codex-external-agent-testskill。它的本机 profile 只配置这两个外部 evaluator,不包含 Codex target;首次使用或缓存被清理后必须重新 回答 runtime 与认证问题。

04
WORKFLOW / MARKDOWN AUTHORED

论文生成联邦知识快照

flowchart LR
    Input["网页 / DOI / PDF"] --> Find["确认规范 PDF"]
    Find --> Prep["逐页核验<br/>转无图片 TeX"]
    Prep --> Ext["提取论文图"]
    Ext --> Ask["查询知识库"]
    Ask --> Gate{"身份状态"}
    Gate -->|known| Bridge["只建立 bridge"]
    Gate -->|new / partial| Entry["解释缺失知识"]
    Gate -->|uncertain / conflict| Review["保留证据"]
    Bridge --> Snap["联邦快照"]
    Entry --> Snap
    Review --> Snap
    Snap -.->|明确要求导入| Put["事务入库"]
    Put --> Brain["个人知识库"]

extract-paper-concepts 先从网页、DOI、标题或 PDF 定位规范全文,保留 PDF hash 与逐页证据,将其规范化成只含文字、原生公式、LaTeX 表格、可信 TikZ/PGFPlots 重建和不可转换图表文字描述的 TeX。预处理通过覆盖、编译与 视觉核验后,才提取候选图并调用 query-kgdistiller。已知 概念不重复解释,只作为论文图与个人图谱之间的 bridge;新知识和论文特有内容留在 独立的联邦快照中。

默认流程不修改个人知识库。只有收到明确导入要求时,才为选中的 new / partial 节点保留论文来源并调用 ingest-kgdistillerknown 仍然只写 ref

03 / OWNED SKILLS

自有 Skill 详情

详情继续按实际目录分类,但保持展开的内容分组,不重复使用文件夹折叠交互。

ROOT

顶层 Skills

skills/ · 03
01 codex-external-agent-testskill

codex-external-agent-testskill

Let Codex atomically evaluate exactly one Agent Skill through isolated Claude Code or OpenCode processes selected from a machine-local runtime profile. Use only when the user explicitly requests an external evaluator, cross-runtime comparison, fresh-process authentication, or process-level isolation. Gate first use on discovering Claude Code and OpenCode, choosing one, and recording subscription or API-file authentication before any fixture, staging, dry-run, or trial work. Do not configure or launch Codex as a target, and do not use for production workflows or ordinary single-Skill tests; codex-subagent-testskill is the default test path.

02 codex-subagent-testskill

codex-subagent-testskill

Default path for evaluating exactly one Agent Skill with fresh native Codex subagents inside the current Codex session. Use for Skill discovery smoke tests, behavioral conformance, negative and safety cases, regression checks, repeated stability trials, and bounded concurrent trials when process-level runtime isolation is not required. Do not use for production deliverables or multi-Skill workflows; use codex-subagent-workflow for production work. Use codex-external-agent-testskill only when the user explicitly requests Claude Code, OpenCode, cross-runtime comparison, or fresh-process authentication.

03 codex-subagent-workflow

codex-subagent-workflow

Execute production work inside the active Codex session with one or more native Codex subagents and optional Agent Skills. Prefer project-defined roles from the standard trusted `.codex/config.toml` `[agents]` configuration; only derive a topology from the request when no custom project roles are configured. Use when research, implementation, review, validation, or other deliverable-producing work benefits from bounded worker roles, parallel investigation, or staged handoffs. Do not launch external agent CLIs, use this Skill for isolated Skill evaluation, or claim process-level isolation; use codex-subagent-testskill to test exactly one Skill.

SUITE

gamemaker/

skills/gamemaker/ · 05
04 gamemaker/build-unity-scene

build-unity-scene

Read a Unity project's architecture and level requirements, then create or modify a Unity scene that respects the documented code boundaries, content structure, asset licenses, and validation workflow. Use when a programming agent needs to lay out a level, wire scene objects and components, create reusable prefabs, bind existing gameplay scripts, add camera and lighting, configure build scenes, or turn a documented prototype into a technically valid Unity scene. This skill builds and validates scenes but leaves exploratory gameplay evaluation to play-unity-game.

05 gamemaker/configure-unity-mcp

configure-unity-mcp

Install, configure, repair, migrate, and fully validate Codex-to-Unity Editor MCP integration on macOS, Windows, or Linux. Use when setting up Unity MCP on a new project or machine; enabling Codex to read Hierarchy and Inspector state, edit scenes and components, control Play Mode, inspect Console logs, or capture Game View; diagnosing missing Unity tools, failed MCP handshakes, package compilation errors, stale Bee/Library caches, port conflicts, multi-instance routing, or Codex configuration scope; or proving that an existing Unity MCP connection works end to end.

06 gamemaker/discuss-game-design

discuss-game-design

Read a game's design documents and discuss a concrete game-design decision while preserving source hierarchy and separating confirmed facts, synthesis, proposals, and open questions. Use when a design agent needs to understand the current vision, analyze a mechanic or loop, compare design options, resolve a design tension, define a prototype, write acceptance criteria, or prepare a clear handoff to art and programming. Discussion is read-only unless the user explicitly asks to update design documents.

07 gamemaker/play-unity-game

play-unity-game

Play and evaluate a Unity game or scene after implementation, using the Unity Editor and the closest available player-like input path. Use when Codex needs to actually play a Unity build or Editor scene, verify a gameplay loop, reproduce a gameplay bug, inspect feel and feedback, capture milestone evidence, check reset/replay and soft-lock behavior, or provide a read-only playtest report for design, art, or programming work. Do not use this skill to implement fixes or build scenes unless the user separately asks for changes.

08 gamemaker/search-game-art

search-game-art

Search the web for game art assets and produce a source- and license-verified shortlist without downloading or importing files. Use when an art agent needs candidate 2D or 3D models, textures, materials, icons, UI packs, animations, VFX, fonts, or environment kits for a game; when a placeholder needs an open or commercially usable replacement; or when the team needs options compared by visual fit, technical fit, license, attribution, format, and scope. This initial skill is search-only.

SUITE

kgdistiller/

skills/kgdistiller/ · 05
09 kgdistiller/extract-and-export-notes

extract-and-export-notes

Extract candidate knowledge graphs from Git-modified Markdown, Typst, and LaTeX notes across mathematical, technical, scientific, humanities, architecture, and everyday domains; preserve user-written knowledge markers; delegate existing-knowledge resolution to query-kgdistiller; delegate reviewed graph writes to ingest-kgdistiller; and publish validated notes to the web. Use when Codex authors, updates, migrates, exports, or validates knowledge-bearing notes at file, directory, course, subject, or repository scope.

10 kgdistiller/extract-paper-concepts

extract-paper-concepts

Find the canonical PDF for a paper URL, DOI, title, or landing page; normalize PDF or other non-TeX input into a visually verified, image-free TeX transcription; extract a source-grounded isolated concept graph; query kgdistiller before writing full entries; omit duplicate entries for knowledge already present in the personal knowledge base; and return a federated paper-to-personal snapshot without mutating the personal graph. Use for web-linked papers, PDFs, scans, LaTeX or Markdown paper repositories, Word manuscripts, prerequisite learning maps, paper concept inventories, and explicit, separately authorized imports of selected missing concepts.

11 kgdistiller/ingest-kgdistiller

ingest-kgdistiller

Apply a reviewed, source-backed knowledge update through kgdistiller's transactional ingest API and return a canonical receipt. Use after query-kgdistiller or another extractor has decided identities, native authority markers, refs, entries, aliases, direct semantic edges, and optional mappings; for changed Markdown, Typst, or LaTeX knowledge; for explicitly authorized paper imports; or for reviewed cross-namespace alignment persistence.

12 kgdistiller/query-kgdistiller

query-kgdistiller

Query a kgdistiller external-brain index without reading or mutating its graph files. Use when an Agent must batch-resolve concept names and aliases, retrieve a bounded source-backed neighborhood, align a candidate note or paper graph, classify concepts as known, partial, new, conflicting, or uncertain, or connect an isolated graph to a personal graph before deciding which entries or refs to author.

13 kgdistiller/trace-concept-lineage

trace-concept-lineage

Batch-distill a technical paper or concept inventory into source-backed concept dossiers, a typed knowledge graph, and a prerequisite-ordered reading route. Collapse generic calculus, linear algebra, and probability foundations into one assumed MATH-FOUNDATION node, and delegate independent domain concepts to parallel subagents that write disjoint files. Use when Codex must turn a paper into a one-session learning map, research many related concepts with authoritative web sources, or prepare non-interactive study material grounded in a target paper.

SUITE

notes/

skills/notes/ · 02
14 notes/create-latex-math-notes

create-latex-math-notes

Create a lightweight LaTeX-authored mathematics notes project under qlblog/notes/math using the synchronized ElegantBook syntax and the repository's LaTeX-to-Typst adapter. Use when the user asks to start, scaffold, initialize, or新建 `.tex` math notes that preview only as Typst HTML, retain `\kn`/`\knref` knowledge markers, and never add PDF, latexmk, SyncTeX, MkDocs, or committed build artifacts.

15 notes/create-math-notes

create-math-notes

Create a new lightweight Typst-first mathematics subject or course under qlblog/notes/math. Use when the user asks to start, scaffold, initialize, or新建 math notes that should open directly in VS Code/Tinymist and inherit the shared QLNotes template, optional ignored LaTeX/Markdown snapshots, semantic HTML checks, and knowledge-graph synchronization.

© 2026 Qiulin. All Rights Reserved. / RSS / Sitemap
Powered by Astro & Fuwari