对齐

工程开发 社区

技能解读

当前原文的自动摘译,尚未人工审校。

协议

调用ck_context加载当前会话状态、域包以及已记录的任何先前目标或决策。致电ck_memory_search,告知用户既定的目标,在提出多余问题之前先展示同一区域先前对齐的工作。一次向用户询问一个问题。对于每个问题,提供您推荐的答案,以便用户可以确认、调整或替换它——永远不要让他们茫然。

不可协商的规则

即使对于“小”任务,也不要跳过对齐并直接进行规划或编码。此阶段错过的约束会导致发现结果受阻或稍后返工。在此技能期间不要制定计划、PRD 或代码。输出是一个记录的目标和一组决策,仅此而已。如果用户说“就做吧”,记录下你到目前为止所知道的,明确记下未知的内容,然后继续前进——但将未解决的问题作为警告显示出来,这样如果他们稍后咬牙切齿,它们就可以作为调查结果浮出水面。

你生产什么

在该技能结束时:ck_goal记录:目标、验收标准、接触层、已知约束、明确假设、开放未知数。每个已解决的设计选择有一个或多个ck_memory_record条目(类型:decision)。关于是否继续进行plan-slice或直接进行ck_review_submit的明确建议。

其他资源

有关完整的受控工作流程,请参阅references/workflow.md

流狐档案 运行与权限由流狐静态检查推断,不代表作者声明或实际测试结果。
领域
工程开发
作者声明的智能体
未找到明确声明;不据此推断已兼容或已测试
静态检查
94 / 100 · 启发式扫描,不代表运行安全
运行环境证据
未执行;当前没有隔离运行测试证据
作者 / 版本 / 许可证
@aryaminus · Apache-2.0
上下文用量评级
低消耗
接入复杂程度
即装即用
是否需要外部接口密钥
不需要
兼容的系统
未声明(默认跨平台)
运行环境要求
未检出明确要求
检测到的文件与系统行为
  • 只读
网络行为
仅限本地
安装命令数
无(仅作为资料)

需要注意: 未声明工具权限范围;实际权限由运行环境决定。

作者说明文件(原文) SKILL.md

# Align Skill

Reach a **shared design concept** with the user before any plan or code is written. The most expensive misalignment is the one caught after implementation — catch it here instead.

## Protocol

1. Call `ck_context` to load current session state, domain pack, and any prior goals or decisions already recorded.
2. Call `ck_memory_search` with the user's stated goal to surface prior aligned work on the same area before asking redundant questions.
3. Ask the user **one question at a time**. For each question, provide your recommended answer so the user can confirm, adjust, or replace it — never leave them staring at a blank.
4. Work through the alignment tree in this order:

   **What** — What is the exact outcome? What is explicitly out of scope?
   **Why** — What problem does this solve? What is the success signal?
   **Who** — Who uses the result? Which roles, systems, or integrations are affected?
   **Layers** — Which system layers does this touch? (schema, services, APIs, UI, infra, third-party) Record each touched layer explicitly — this drives vertical slice decomposition later.
   **Acceptance criteria** — What does "done" look like? What would a failing test catch? Reframe vague requirements into concrete, testable conditions.
   **Edge cases** — What breaks if inputs are invalid, empty, or unexpected?
   **Constraints** — Budget, timeline, tech stack limits, compliance requirements, or must-not-change areas.
   **Assumptions** — What are you assuming about the environment, tech stack, data sources, or integrations? Surface these explicitly before proceeding.
   **Unknowns** — What do you not know yet? What needs a spike or research before implementation starts?

5. After each resolved decision, call `ck_memory_record` with type `decision` to persist it. Future agents resuming this work will recover these without asking again.
6. When alignment is complete, call `ck_goal` (mode: `record`, horizon: `session`) to record the aligned goal with its acceptance criteria and touched layers.
7. Tell the user the alignment is complete and recommend the next step:
   - If the work is multi-layer and will need decomposition → activate `plan-slice`.
   - If the work is a single small change that fits in one task → proceed to `ck_review_submit` (plan_phase: `narrowed_decision`).

## Non-negotiable rules

- Never skip alignment and go straight to planning or code, even for "small" tasks. A missed constraint at this stage compounds into blocked findings or rework later.
- Do not produce a plan, PRD, or code during this skill. The output is a recorded goal and a set of decisions — nothing else.
- If the user says "just do it," record what you know so far, note the unknowns explicitly, and move forward — but surface the open questions as warnings so they can surface as findings if they bite later.
- Planning is always human-in-the-loop. The agent asks; the human decides. Never answer your own alignment questions and continue as if the human agreed.

## What you produce

At the end of this skill:
- A `ck_goal` record with: objective, acceptance criteria, touched layers, known constraints, explicit assumptions, open unknowns.
- One or more `ck_memory_record` entries (type: `decision`) for each resolved design choice.
- A clear recommendation on whether to proceed to `plan-slice` or directly to `ck_review_submit`.

## Additional resources

- For the full governed workflow, see [references/workflow.md](references/workflow.md)

讨论

基于 GitHub Discussions。登录 GitHub 即可参与讨论、点赞、订阅更新。