| name | executing-plans |
| description | Use when a written plan exists and you are about to build it in the current session, to work it task by task with a checkpoint after each. Keeps execution honest against the plan instead of drifting into a freehand build. |
| slipstream | {"category":"context","requires":[],"tags":["planning","execution","verification"]} |
Overview
A plan is only worth writing if it is followed. The failure mode is opening the plan, reading the goal, and then building from memory — the plan becomes decoration and the checkpoints never happen. This skill is the discipline of executing a plan as written: one task at a time, verify before moving on, and surface any divergence rather than quietly absorbing it.
It is the counterpart to [[write-plan]]: that skill produces the plan; this one spends it. For independent tasks you may hand each to a fresh agent ([[subagent-driven-development]]); this skill is the in-session path where you build the tasks yourself.
Steps
- Open the plan and work the tasks in order. Do not skim ahead and build the whole thing; the order encodes dependencies.
- For each task, do exactly what it specifies, then run its stated verification before starting the next. A task is not done until its check is green.
- Commit at each task boundary so the history matches the plan and a failed task is cheap to undo.
- When reality contradicts the plan — a step is wrong, missing, or impossible — stop and fix the plan first, then continue. Never silently improvise past a divergence; the plan must stay the source of truth.
- After the last task, re-run the full suite and the plan's overall acceptance criteria together, not just the final task's check.
Verify
Every task was built and verified in plan order, each task boundary is a commit, any divergence was written back into the plan before continuing, and the full acceptance criteria pass at the end.
