AgentSkillsCN

ap-init

通过创建持久化的“真相文件”(PROJECT.md、DECISIONS.md、PROGRESS.md、BACKLOG.md、STATUS.md),对项目进行初始化或重新梳理。适用于新项目启动之时,或当项目范围发生偏离时使用。

SKILL.md
--- frontmatter
name: ap-init
description: Initialize or re-ground a project by creating durable truth files (PROJECT.md, DECISIONS.md, PROGRESS.md, BACKLOG.md, STATUS.md). Use when starting new work or when scope has drifted.

/ap:init

Purpose

Initialize (or re-ground) a project using agent-pack.

This command creates the durable truth set and the initial working surface:

  • PROJECT.md (what we're building)
  • DECISIONS.md (durable decisions)
  • PROGRESS.md (completed work and learnings)
  • BACKLOG.md (problem list)
  • STATUS.md (what's happening next)

Use /ap:init when starting a new repo or when scope has drifted.


Inputs

Provide as much as you can. Rough notes are fine.

  • Project idea / description
  • Target users
  • Constraints (time, budget, tech, compliance)
  • Preferred stack (optional)
  • Non-goals (what we are not doing)

If inputs are missing, this command must surface them as Unknowns / needs input.


Reads

Must consult the following if they exist:

  • .agent-pack/context/PROJECT.md
  • .agent-pack/context/DECISIONS.md
  • .agent-pack/context/PROGRESS.md
  • .agent-pack/work/BACKLOG.md
  • .agent-pack/work/STATUS.md

If these do not exist, create them from templates.


Process

  1. Clarify the one-liner
    • Produce a one-sentence description that can be repeated later.
  2. Extract scope shape
    • Identify users + top use cases
    • Identify non-goals
    • Identify constraints
  3. Define success criteria
    • Produce 2-6 checkable success criteria.
  4. Seed decisions
    • Record any obvious durable decisions as Locked.
    • Record unresolved forks as Pending.
  5. Seed the backlog
    • Produce 3-10 problem statements.
    • Keep them problem-first, not solution-first.
  6. Set immediate next actions
    • Update STATUS.md with current focus + next 3 actions.
    • Prefer actions that lead naturally to /ap:plan.

Output rules

  • Follow OUTPUT CONTRACTS.
  • Keep content short and readable.
  • Do not invent requirements.
  • If uncertain, explicitly mark assumptions.

Output contract

md

### Files updated

- .agent-pack/context/PROJECT.md -- <summary>
- .agent-pack/context/DECISIONS.md -- <summary>
- .agent-pack/context/PROGRESS.md -- <summary>
- .agent-pack/work/BACKLOG.md -- <summary>
- .agent-pack/work/STATUS.md -- <summary>

### Decisions made

- <D/P entries>

### Tasks created or modified

- (usually none; /ap:init seeds backlog, not plan tasks)

### Checks created or affected

- (none; success criteria live in PROJECT.md)

### Unknowns / needs input

- <questions>


Completion criteria

/ap:init is complete when:

  • PROJECT.md has a clear one-liner, users, non-goals, constraints, and success criteria
  • DECISIONS.md contains at least one item (Locked or Pending)
  • PROGRESS.md exists (may contain only the template)
  • BACKLOG.md has at least 3 meaningful problem statements
  • STATUS.md has a current focus and 3 next actions

If any of the above are missing, the command must report it.


Common failure modes

  • Writing tasks instead of problems (belongs in PLAN.md)
  • Missing non-goals (scope creep risk)
  • Vague success criteria ("easy", "fast", "robust" without signals)
  • Silent assumptions instead of explicit unknowns

Notes

/ap:init should feel fast.

If this command becomes long or painful, it is a signal that the project definition is unclear -- not that more process is needed.