Source-first project skeletons for LLM-assisted development.
dev-skeleton ships short, LLM-readable skill prompts and copyable templates. It is not tied to any specific agent runtime, and it does not ship a CLI, code indexer, KB builder, or project-management workflow.
Skeletons orient the model before it reads source. They do not answer implementation questions.
- Source is authority: code, config, tests, release artifacts, and maintained docs.
- Skeleton is orientation: purpose, non-goals, stable constraints, truth sources, and review preferences.
- Do not mirror modules, classes, functions, call graphs, APIs, or current behavior.
- Let the model inspect source for the task at hand.
- Update skeletons only when durable intent, constraints, truth sources, or review preferences change.
- Routine implementation changes should not update skeletons.
Source-first is economical only when the code tree preserves clear semantic ownership. Code that shares one intent, lifecycle, invariant, or failure policy should stay together; responsibilities that change independently should separate. Large cohesive modules are acceptable. Mixed-responsibility hubs are not.
Root documentation should route readers to subsystems, and entry modules should route them to semantic owners. Implementation facts still belong in source. Dev-skeleton records durable navigation constraints, not a second inventory of the codebase.
skills/
skeleton-init/
skeleton-refresh/
skeleton-audit/
review-skeleton/
templates/
DEV_SKELETON.md
REVIEW_SKELETON.md
AGENTS.md
Root DEV_SKELETON.md, REVIEW_SKELETON.md, and AGENTS.md describe this repo itself.
Copy the templates into a target repo, then give the relevant skill prompt to your LLM or agent environment:
skeleton-init: create initial skeleton files.skeleton-refresh: update skeletons after durable direction changes.skeleton-audit: find over-detail, stale claims, and source-first violations.review-skeleton: review source and diffs using project review preferences.
Skill distribution is intentionally external to this repo.
- No persistent implementation KB.
- No source index.
- No CLI.
- No test, onboarding, release, or multi-agent workflow.
When a skeleton starts explaining how the code works today, delete that detail and point back to source.
面向 LLM 辅助开发的 source-first 项目骨架。
dev-skeleton 提供简短、可被 LLM 直接阅读的 skill prompt 和可复制模板。它不绑定任何特定 agent runtime,也不提供 CLI、代码索引、KB 构建器或项目管理流程。
Skeleton 的作用是在模型阅读源码前提供方向感。它不回答实现细节问题。
- 源码优先:代码、配置、测试、release 产物和维护中的文档才是权威。
- 骨架只做定向:记录目的、非目标、稳定约束、事实来源和 review 偏好。
- 不镜像模块、类、函数、调用图、API 或当前行为。
- 让模型根据当前任务动态阅读源码。
- 只有长期意图、约束、事实来源或 review 偏好变化时,才更新 skeleton。
- 日常实现改动不应该触发 skeleton 更新。
只有当代码树保持清晰的语义所有权时,source-first 才是经济的。共享同一意图、生命周期、不变量或失败策略的代码应当聚合;因不同原因独立变化的责任应当分离。大而内聚的模块可以保留,混合责任的中心文件不应继续增长。
根文档负责指向子系统,入口模块负责指向语义 owner,实现事实仍以源码为准。Dev-skeleton 只记录长期有效的导航约束,不维护第二份代码清单。
skills/
skeleton-init/
skeleton-refresh/
skeleton-audit/
review-skeleton/
templates/
DEV_SKELETON.md
REVIEW_SKELETON.md
AGENTS.md
根目录的 DEV_SKELETON.md、REVIEW_SKELETON.md 和 AGENTS.md 描述的是本仓库自身。
把模板复制到目标仓库,然后把对应的 skill prompt 交给你使用的 LLM 或 agent 环境:
skeleton-init:创建初始 skeleton 文件。skeleton-refresh:在长期方向变化后更新 skeleton。skeleton-audit:检查过度细节、陈旧声明和 source-first 违规。review-skeleton:基于项目 review 偏好审查源码和 diff。
Skill 分发方式刻意留给外部系统处理。
- 不做持久化实现 KB。
- 不做源码索引。
- 不做 CLI。
- 不做测试、新人引导、release 或多 agent 工作流。
当 skeleton 开始解释“当前代码如何工作”时,删掉这部分细节,让模型回到源码。
