AgentSkillsCN

ap-do

从当前活动计划中精确执行一项任务,更新任务状态并记录相关证据。适用于当某项已明确检查点的单一任务即将进入构建阶段时使用。

SKILL.md
--- frontmatter
name: ap-do
description: Execute exactly one task from the active plan, update status, and record evidence. Use when a single task with a defined check is ready to build.

/ap:do

Purpose

Execute one task from the current plan.

/ap:do is intentionally narrow: it builds a single task, updates status, and records what happened. If more than one task is executed, the command has failed.


Inputs

Required:

  • A task ID from PLAN.md (e.g. T001)

Optional:

  • Notes or constraints (e.g. "prioritize correctness over speed")

If no task ID is provided, pick the next incomplete task from the PLAN.md.


Reads

Must consult:

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

May consult:

  • Repo source code
  • Relevant runs/ history

Process

  1. Validate the task
    • Confirm the task exists and is in the active plan.
    • Confirm the task has exactly one associated check.
  2. Reconstruct intent
    • Re-read task scope, done-when condition, and related decisions.
    • If intent is unclear, stop and report an unknown.
  3. Execute the task
    • Make the minimum changes required to satisfy the task.
    • Do not perform unrelated refactors or improvements.
  4. Self-check against the defined check
    • Run or simulate the check.
    • Do not invent new checks during execution.
  5. Record the run
    • Create a new folder under runs/:
    • <timestamp>_<task-id>/
    • Record inputs, outputs, and notes.
  6. Update status
    • Mark progress in STATUS.md.
    • Update "Recent changes" with a pointer to the run.
  7. Append progress log
    • Add a dated entry to PROGRESS.md with a short summary and any learnings.

Output rules

  • Follow OUTPUT CONTRACTS.
  • Change only what the task requires.
  • Prefer correctness and clarity over cleverness.
  • If the task reveals a missing decision, stop and report it.

Output contract

md
### Files updated

- <file path> -- <what changed>
- .agent-pack/context/PROGRESS.md -- <progress entry>

### Decisions made

- (none expected; new durable decisions should be rare)

### Tasks created or modified

- T### -- <status update or note>

### Checks created or affected

- C### -- <result or notes>

### Unknowns / needs input

- <questions or blockers>

Completion criteria

/ap:do is complete when:

  • The task's done-when condition is satisfied
  • The associated check can be run
  • STATUS.md reflects the new state
  • A run record exists under runs/
  • PROGRESS.md records the completed work

If the task cannot be completed, the command must report why.


Common failure modes

  • Executing multiple tasks in one run
  • Expanding scope beyond the task definition
  • Quietly changing decisions
  • Creating new checks mid-execution
  • Forgetting to update STATUS.md

Notes

/ap:do should feel focused and boring.

Momentum comes from finishing small tasks, not heroic executions.