What it looks like
Task: move memos from the old server to a new one, with the data, with no more than 10 minutes of downtime.
Don't do anything yet. First:
1. Ask what's missing for a plan — no more than five questions, the most important ones.
2. List the assumptions you'll make if there's no answer.
Then, after the answers, a step-by-step plan with a check and a rollback at each step.
The model’s typical questions here: where the data lives, which database, whether there’s a backup, which domain and where DNS is, whether the new server can be brought up in parallel. It would otherwise decide each of them for you.
Why
- The model is trained to be helpful right away: without permission to ask, it doesn’t ask, it guesses.
- A plan is cheaper than code. A mistake in a plan is fixed with one sentence; a mistake across twenty changed files, with a rollback.
- An agreed plan is a done criterion: you check against it that everything was done (Where models fail).
In agents
Claude Code has a plan mode: the agent reads files and writes a plan but changes nothing until the plan is approved. For a task spanning several files that’s the first step; for a one-line edit it’s overkill. The same principle is in Running a server with an AI agent: the agent proposes, the human decides.