菜单
Break a large or multi-session engineering objective into ordered, verifiable stages with dependencies and checkpoints. Use when the work spans several components or cannot be safely held in one execution pass. This skill plans but does not implement.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/dayfinggg/openai-codex-agent-skills --skill roadmap命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
Design types, interfaces, module boundaries, and ownership before implementing a non-trivial feature or structural change. Use when an early design choice would constrain several files or components. Do not use for a small local edit or when an approved design already exists.
Prepare and assign one bounded, independently verifiable subtask to a subagent. Use when delegation is available and authorized, one work unit can proceed without shared mutable state, and parallel or specialized work will materially help. Do not use merely to avoid doing a small task locally.
Reproduce a failure, regression, or performance problem and identify its root cause with evidence. Use when the user asks why something is broken, failing, crashing, flaky, or slow. Do not modify the implementation unless the request also asks for a fix.
正在显示 SKILL.md
| name | roadmap |
| description | Break a large or multi-session engineering objective into ordered, verifiable stages with dependencies and checkpoints. Use when the work spans several components or cannot be safely held in one execution pass. This skill plans but does not implement. |
Create a route to the outcome that remains usable across sessions and agents.
State the final observable outcome, current state, constraints, non-goals, and definition of done. Resolve only ambiguities that change sequencing or architecture.
Split work into vertical units that leave the system in a valid, testable state. Make dependencies explicit. Put risk-reducing discovery and irreversible decisions early. Avoid tasks titled with multiple independent outcomes.
Before a stage creates shared infrastructure, a generic framework, or a cross-cutting abstraction, require evidence from one working vertical scenario and multiple real consumers. Make generalization a checkpoint decision rather than an assumption in the first stage.
Each unit must state its result, likely scope, prerequisites, acceptance criteria, and concrete verification. Identify safe parallel work only after contracts and shared ownership are settled.
Limit simultaneous in-progress stages to the work the available owners can finish and integrate. Starting more lanes is not progress when reviews, dependencies, or verification are already the bottleneck.
For date-sensitive work, record an optimistic, nominal, and pessimistic range with its assumptions, dependency owners, next checkpoint, and fallback trigger. Do not present a nominal estimate as a promise. Update the forecast when evidence changes instead of preserving a stale date.
Ensure the sequence contains no circular dependency, hidden migration step, unverified handoff, or stage that is too large for one focused session. Add checkpoints where evidence should be reviewed before more work depends on it.
Do not start implementation. Do not overwrite an existing incomplete plan for different work. Do not create external tickets without authorization.
Return the destination, assumptions, ordered stages, dependency edges, checkpoints, risks, and open decisions.
