Executing Plans Skill
An approved plan is a contract. Execute exactly what was agreed — nothing more, nothing less.
This skill is the execution phase of the plan-first → execute workflow.
Use plan-first to write and get approval; use this skill to run it.
When to use this skill
- User approved a plan and says "go", "proceed", "do it", "execute", "run it"
- You have a written plan in context with numbered steps
- Mid-plan: before each step, to stay on track
Execution rules
One step at a time
Run each plan step independently. Do not bundle step 2 and 3 into one edit "to save time." Show what changed after each step.
No scope expansion
If you notice a related improvement mid-execution:
- Do NOT add it to the current step
- Write a note: "Noticed: [X]. Not in scope — raise after this plan completes."
- Continue with the next plan step
Stop conditions
Pause and ask the user if any of:
- A step reveals a file/system state different from what the plan assumed
- A step requires a decision that was not in the plan
- A step fails and there is more than one plausible fix
Do not guess. Do not proceed past a blocker silently.
After each step
Report briefly:
Step N complete.
Changed: [file or command output]
Next: Step N+1 — [description from plan]
After all steps
Run the plan's "Verify after" commands and show output verbatim.
Then use the verify-before-done skill before claiming the plan is complete.
What NOT to do
- Do not skip a step because it "seems fine"
- Do not reorder steps
- Do not add an unplanned commit, lint run, or refactor
- Do not mark the plan done before verification passes