AgentSkillsCN

executing-plans

在可控批次中分阶段执行实施方案,并在每批次之间设置检查点,进行验证,同时收集用户反馈。

SKILL.md
--- frontmatter
name: executing-plans
description: Execute implementation plans in controlled batches with checkpoints, verification, and user feedback between each batch.
autoApply: true

Executing Plans

Announce at start: "I'm using the executing-plans skill to implement this plan."

Purpose

Execute an existing implementation plan methodically -- in batches, with verification at every step, and with explicit user checkpoints between batches.

Branch Safety

Never start implementation on main or master without explicit user consent. If the current branch is main or master, ask the user to confirm or specify a feature branch before proceeding.

Execution Protocol

Step 1: Load and Review Plan

  1. Read the plan in its entirety.
  2. Identify concerns before writing any code:
    • Are there ambiguous steps?
    • Are there missing dependencies or prerequisites?
    • Are there tasks that conflict with the current codebase?
    • Are file paths still accurate?
  3. Raise issues with the user before starting. Do not silently work around gaps.

If the plan is sound, confirm: "Plan reviewed. No blocking issues found. Beginning execution."

Step 2: Execute Batch

Execute tasks in batches. Default batch size is 3 tasks unless the user specifies otherwise.

For each task in the batch:

  1. Mark the task as in_progress in the todo list.
  2. Follow the plan steps exactly. Do not improvise, skip steps, or reorder.
  3. Run all verification commands specified in the plan.
  4. Mark the task as completed only when verification passes.

If a task's verification fails:

  • Stop the batch.
  • Report the failure with full output.
  • Wait for user guidance.

Step 3: Report

After each batch, present a checkpoint report:

code
## Batch Report

### Completed
- Task 1: <description> -- PASSED
- Task 2: <description> -- PASSED
- Task 3: <description> -- PASSED

### Verification Output
<paste relevant test output, build output, or command results>

### Issues Encountered
<any warnings, deviations, or concerns>

End every report with: "Ready for feedback."

Wait for the user to respond before continuing.

Step 4: Continue

  1. Apply any feedback the user provides (code changes, plan adjustments, re-ordering).
  2. Execute the next batch following the same protocol.
  3. Repeat until all tasks are complete.

Step 5: Complete Development

When all tasks are done:

  1. Run the full test suite and report results.
  2. Run linters and formatters if configured.
  3. Present final options:
    • Create a pull request
    • Run additional integration tests
    • Continue with the next plan

Mandatory Stop Conditions

STOP executing immediately when any of these occur:

  • Hit a blocker -- a dependency is missing, a service is down, or a required file does not exist.
  • Plan has critical gaps -- a step is too vague to execute without guessing.
  • Instruction is unclear -- you are unsure what the plan intends.
  • Verification fails repeatedly -- the same test fails after two honest attempts.
  • Scope creep -- a task requires changes far beyond what the plan describes.

When stopped, explain why and ask for guidance. Do not attempt workarounds without user approval.

Principles

  • Fidelity to the plan. The plan is the source of truth. If the plan is wrong, fix the plan first.
  • Transparency. Show verification output. Never summarize away failures.
  • User control. The user decides when to continue, pivot, or stop.
  • Small batches. Frequent checkpoints catch problems early.

Adapted from obra/superpowers