| name | plan-then-act |
| description | Have the agent propose a concrete, numbered plan and get approval BEFORE executing — then work the plan step by step, re-planning out loud when reality diverges. Part of the "keep the agent on track" suite. Use for any task big enough that a wrong first move is costly, or when an agent tends to charge ahead. |
Plan, then act
Why
The cheapest course-correction happens before the work starts. Making the agent commit to a plan first turns a silent wrong turn into a visible, editable proposal.
Structure
# Objective
One line: what we're doing.
# Plan
1. <step> — touches <files/artifacts> — done when <observable result>
2. …
# Resolve first
Unknowns or risks to settle before step 1.
Protocol
- The agent writes the plan and waits for approval (or edits) before touching anything.
- It works one step at a time, reporting what changed vs. what it expected.
- When reality diverges from the plan, it stops and re-plans out loud — it does not improvise past the agreed plan.
Anti-patterns
- A plan so vague it couldn't be wrong ("1. build it. 2. test it").
- Executing before approval.
- Silently deviating from the approved plan when something surprises it.
