Plan-Then-Act

Plan-Then-Act is a prompt pattern for work that has risk, ambiguity, or multiple steps. It asks the agent to state the intended approach before making changes, then execute against that plan with evidence and verification.

Use it when the task touches code, content generation, routes, publishing, assets, or visual layout. The plan should name the target files or systems, the expected output, the constraints, and the checks that will prove the work is complete.

The pattern should not become a long planning ritual. For small visual adjustments, the plan can be one or two sentences. For larger workflow changes, it should list the sequence of inspect, edit, generate, verify, and report.

A useful prompt structure is: state the goal, inspect the current state, identify the smallest safe change, make the change, verify it, and report what changed. If evidence contradicts the plan, update the plan before continuing.

The strongest version of this pattern includes stop conditions. The agent should pause when required data is missing, when the requested change conflicts with project rules, or when the current state differs from the assumptions in the prompt.

Plan-Then-Act works because it makes intent visible. Humans can catch bad assumptions early, and agents can keep scope under control while still moving the task forward.