executing-plans

Executing Plans

Overview

Load the plan, review it critically, execute tasks in order, verify each step, and hand off to the finishing skill. The plan is the source of truth — if you find yourself improvising, stop.

This skill is the inline executor. If your environment supports subagents, prefer workflow-subagent-driven instead — it produces higher-quality output for loosely-coupled tasks.

When to Use

• Plan exists at .agents/plans/YYYY-MM-DD-<name>.md
• Subagents unavailable, or tasks are tightly coupled (each task depends on in-memory context from the previous)
• You want human checkpoints between batches rather than between every task

When NOT to use:

• No plan exists → use workflow-writing-plans first
• Tasks are independent and subagents are available → use workflow-subagent-driven
• Single-file trivial change → just do it, no plan needed

The Workflow

Load plan ─→ Critical review ─→ Create TodoWrite ─→ Execute task loop ─→ Finishing skill
                    │                                      │
                    ▼                                      ▼
             Raise concerns                     Per task: follow steps
             before starting                    exactly, run verifications,
                                                mark complete

Step 1 — Load and Review the Plan

1. Read the plan file end to end.
2. Review critically. Before executing anything, ask yourself:
   • Does every task have the code, verification command, and commit written out?
   • Are type/name usages consistent across tasks?
   • Any placeholders (TBD, TODO, similar to, appropriate)?
   • Does the plan violate universal rules (soft delete, i18n, @latest, Route→Service→DB)?
3. If you find problems → stop and raise them with the human before starting. Don't patch a broken plan silently.
4. If the plan is clean → create a TodoWrite with one entry per task, mark the first as in_progress, continue.