DBX Skills

个人 Claude/Codex Skills 仓库，用于集中管理和版本控制。 稳定技能统一使用 dbx- 前缀，避免和外部安装的 skills 产生命名歧义。

Personal Claude/Codex Skills repository for centralized management and version control. Stable skills use the dbx- prefix to avoid naming conflicts with third-party skills.

Stable Skills

Skill	描述 / Description
dbx-open-source-commit-pr	面向开源仓库的英文 commit/PR 写作。Generate English commit messages and PR descriptions for open-source repositories.
dbx-work-commit-pr	工作场景的中文合同式 commit/PR 写作。Contract-style commit messages and PR descriptions for work contexts.
dbx-write	中文为主的观点写作、技术博客、个人随笔、Markdown 草稿、改稿和英文转写。Chinese-first viewpoint writing for technical blogs, personal essays, Markdown drafts, editing, and occasional English versions.
dbx-read	来源约束阅读：对 URL、PDF、论文、文档、GitHub、本地文件或粘贴文本做抓取证据化、摘要、精读、比较、Markdown 提取和可选笔记沉淀。Source-grounded reading for explicit source material.
dbx-crystallize	模糊产品/软件想法、需求草稿、issue 或前置讨论的需求结晶：澄清用户/场景/状态变化、scope、non-goals、验收标准、边界状态、开放决策和 handoff。Requirement crystallization before product judgment, design, technical planning, or implementation.
dbx-diff-review	精确选择 PR/diff/staged/commit/file review 范围，并输出高信号风险发现。Scoped code-change review for concrete diffs.
dbx-linus-review	严格、实用主义、证据驱动的技术方案、模型和合并风险判断。Strict pragmatic technical judgment.
dbx-technical-plan	代码变更前的证据边界技术实施计划，明确 source of truth、不变量、实施切片、验证模型和 handoff。Evidence-bound technical implementation planning before coding.
dbx-code-ratchet	可修改代码的有界 review-repair-revalidation 元技能：调度 diff/Linus review，triage findings，局部自动修复，验证并在方向错误或风险发散时停止。Bounded code ratchet for concrete diffs.
dbx-architecture-health	仓库/模块架构健康体检：识别长期腐化、状态 owner 混乱、验证拓扑缺口和 AI coding 可操作性风险，并给出只读防腐路线图。Read-only architecture health audits for repo decay and AI-coding operability.
dbx-skill-architect	场景优先的 skill 创建、评审、改进与评测。Scenario-first skill architecture, critique, improvement, and eval design.
dbx-skill-portfolio-auditor	已安装技能组合审计与全局/项目/显式触发/禁用/卸载范围决策。Audit installed skill portfolios and recommend global, project, explicit-only, disable, or uninstall placement.
dbx-product-judgment	证据边界内判断产品、功能、PRD、交互、信息架构、实现对齐、路线图或竞品定位是否产品正确。Evidence-bounded product judgment across artifacts and decisions.
dbx-design-judgment	证据边界内评审 UI、流程、截图、PRD、原型、设计系统或代码支撑的界面设计；只做设计判断和实现交接，不改代码。Evidence-bounded design judgment and handoff without implementation.
dbx-attention-routing	混合输入的注意力分流内核：把收藏、想法、任务、课程、工具、消息等路由为行动、构建、实验、追踪、存档、孵化、丢弃、风险保护或澄清，并支持个人配置和外部系统 dry-run 映射。Product-agnostic attention router for mixed inputs with profile and adapter layers.
dbx-learn	持久学习控制器：用于真正理解概念、资料研究、练习 rep、主动复习和可选学习记录；不替代普通总结、直接编码或混合 inbox 分流。Durable learning controller for capability-building study.
dbx-conversation-align	对话卡点诊断、措辞改写与边界沟通。Diagnose stuck conversations, rewrite risky messages, and plan boundaries.
dbx-decision-framing	高影响真实决策的分支门禁、取舍分析与验证动作。Frame high-impact decisions with gates, trade-offs, and validation steps.
dbx-subagent-context	Codex subagent 上下文继承策略。Set Codex subagent context inheritance strategy.
dbx-goal-writer	Codex goal 合同生成、启动与审计。Create, start, and audit Codex goal contracts.
dbx-agent-handoff	AI agent/session 续跑交接包生成。Create restart handoffs for future AI sessions.
dbx-software-plan-first-plan-issue	手动触发的软件 Plan-First 对话收敛阶段：不读仓库、不写计划文件、不实现代码。Manual-only plan convergence phase.
dbx-software-plan-first-ground-plan	手动触发的软件 Plan-First 只读 grounding 阶段：确认仓库事实、规则、source of truth 和验证模型。Manual-only read-only grounding phase.
dbx-software-plan-first-finalize-plan	手动触发的软件 Plan-First 定稿阶段：写入中文 plan.md / tasks.md 并建立 workflow seal。Manual-only plan finalization phase.
dbx-software-plan-first-implement-feature	手动触发的软件 Plan-First 单任务实现阶段：只做第一个未完成 task，验证后等待 review。Manual-only review-gated implementation phase.
dbx-software-plan-first-showhand	手动触发的软件 Plan-First 安全自动执行阶段：仅在所有门禁满足时连续执行完整计划。Manual-only safe automation phase.

Repository Governance

This repository is not just a prompt collection. It is a set of reusable, evaluable, and evolvable agent work units.

DBX applies Agent Skill Control Theory as a practical skill-collection discipline. ASCT itself lives in the theory repository; this repository keeps only applied rules, routing policy, validation practice, and runtime skills.

Start here when creating or changing a skill:

• DBX_SKILL_STYLE_GUIDE.md: operational writing, architecture, eval, placement, and release rules.
• DBX_SKILL_INDEX.md: current skills, boundaries, maturity, routing notes, and next improvements.
• docs/DBX_ASCT_ADOPTION.md: how this repo applies ASCT 0.3 without duplicating the theory repo.
• docs/DBX_PLACEMENT_GUIDE.md: how to decide whether a control belongs in a skill, script, reference, command, hook, repo memory, or collection routing.
• docs/DBX_COLLECTION_DESIGN.md: collection-level routing, skill graph, conflicts, installation scope, safety, and deprecation.
• docs/DBX_ROUTING_MATRIX.md: conflict resolution and chaining rules across DBX skills.
• docs/DBX_EVAL_GUIDE.md: trigger, process, output, safety, regression, and collection-level evals.
• docs/DBX_STATEFUL_SKILLS.md: state contracts for project memory, bootstrap, workflow state, and interaction modes.
• docs/DBX_HOST_ARTIFACTS.md: commands, hooks, AGENTS.md, CLAUDE.md, llms.txt, status lines, planning files, and portability rules.
• docs/DBX_CODEX_COMPATIBILITY.md: Codex feature drift and compatibility policy.
• docs/DBX_RELEASE_CHECKLIST.md: release checklist for adding or changing skills.
• SECURITY.md: script, dependency, state, and third-party skill safety policy.

Local Checks

# Check frontmatter, directory structure, eval JSON, README/index mentions,
# and plan-first shared-file sync.
python3 scripts/validate_skills.py --root .

# Generate Markdown or JSON inventory.
python3 scripts/skill_inventory.py --root . --format markdown
python3 scripts/skill_inventory.py --root . --format json

# Validate evals/triggers.json files.
python3 scripts/run_trigger_evals.py --root . --validate-only

Strict mode is useful before release, but do not enable it in CI until warnings are intentionally cleaned up:

python3 scripts/validate_skills.py --root . --fail-on-warnings

Usage

Install with add-skill

# Install all skills.
npx add-skill dbvc/skills

# Install one skill.
npx add-skill dbvc/skills --skill dbx-open-source-commit-pr

# List available skills.
npx add-skill dbvc/skills --list

Manual install

git clone https://github.com/dbvc/skills.git
cd skills

# Codex / Agent Skills current user scope
mkdir -p ~/.agents/skills
for skill in skills/dbx-*; do
  ln -sf "$(pwd)/$skill" ~/.agents/skills/
done

# Legacy or custom Codex setups
mkdir -p ~/.codex/skills
for skill in skills/dbx-*; do
  ln -sf "$(pwd)/$skill" ~/.codex/skills/
done

# Cursor
mkdir -p ~/.cursor/skills
for skill in skills/dbx-*; do
  ln -sf "$(pwd)/$skill" ~/.cursor/skills/
done

Use ~/.agents/skills for the current Codex / Agent Skills user scope. Keep ~/.codex/skills only for legacy or custom Codex setups that still discover skills there.

Creating a New Skill

Do not copy an existing skill and start stacking rules. Start with the scenario and failure map:

Scenario name:
Primary user:
Task distribution:
Typical inputs:
Expected outputs:
Base-agent failure modes:
Evidence sources:
Hard constraints:
Non-goals:
Success criteria:
Cost and risk budget:

Minimum expectations:

1. skills/<name>/SKILL.md has valid YAML frontmatter and name matches the directory.
2. description explains capability and trigger context, including near-miss boundaries.
3. Full skills have evals/triggers.json and evals/evals.json, or a written reason for deferring one.
4. Fragile, repeatable, mechanical, or checkable work belongs in scripts/, not prose.
5. Long knowledge, rubrics, examples, and gotchas belong in references/, not in a swollen SKILL.md.
6. Before adding a control, decide placement: skill, reference, script, asset, repo memory, command, hook, collection routing, or global instruction.

Related Resources

• Agent Skill Control Theory
• Agent Skills Specification
• Agent Skills Evaluation Guide
• Agent Skills Scripts Guide
• Anthropic Skills Repository
• OpenAI Skills Repository

License

MIT