1.4 Planning mode: agree the plan first

The failure that wastes the most time is not a bad answer. It is a good answer to a question you did not ask, arrived at after nine minutes of work, in a shape that is hard to salvage.

Planning mode fixes it for the cost of one exchange. Before executing, the agent produces a plan — an ordered list of what it intends to do, which files it will read, and what it will produce. You approve it, edit it, or throw it away.

Ask for the plan even when it is not offered

The interface has a planning affordance, and the label has moved between builds. It does not matter: you can always get the same thing by asking.

Before you write anything, give me a plan: the steps in order, which files you will read for each, what you will produce, and the two places you are most likely to get this wrong. Wait for my approval.

That last clause — the two places you are most likely to get this wrong — is worth more than the rest of the plan. It surfaces the assumption the agent is about to make silently, which is usually the thing that ruins the output.

What a bad plan looks like

Reject the plan when you see these
Signal	What it means
No file is named in any step	It is going to write from general knowledge. Your company will not appear in the output except by name.
Step one is “understand the requirements”	Filler. Ask what it will actually read.
Every step is the same size	It has not thought about the work, it has listed the shape of a plan.
It plans to “research current best practice”	Unless it can browse and you want that, this is where invented citations come from.
Nothing is listed as risky	Ask again. There is always something.

Exercise

Ask for a plan for a genuinely fiddly piece of work: “Turn @research/interviews-march.md into a one-page activation problem statement for the leadership review.”
Read the plan. Change exactly one thing — the audience, the length, or the file it reads — and re-approve.
Compare the result against what you would have got from the bare request. Note how much of your one edit shows up in the output.

When to skip it

Planning is overhead on short work. Rewriting a paragraph, tightening a sentence, naming a feature — just ask. The rule of thumb: if the task would take you more than about fifteen minutes by hand, get a plan. If not, don’t.