Ask for the plan before the work

The cheapest correction is the one before the work

You ask for a 2,000-word article, a migration script, a full analysis. Four minutes later you have something substantial built on an assumption you would have rejected in five seconds.

Ask for the plan first.

Before writing anything: give me the outline as six headings with one line each on what goes under them, and list the three assumptions you are making about the audience and scope. Do not write the article yet.

You read fifteen lines. You correct the two that are wrong. Then you say go.

The cost of correcting an assumption rises steeply with how much has been built on it, and this is the move that catches it at the bottom of that curve.

Ask for the assumptions explicitly

The outline is useful. The assumptions list is more useful, and almost nobody asks for it.

Every substantial request leaves things unstated, and the model fills them silently: who the reader is, what the timeframe is, whether the figures are annual or monthly, whether “customers” means accounts or people. A plan surfaces those as a list you can scan.

List the assumptions you are making that, if wrong, would change the result.

That phrasing is deliberate. It filters out trivia and produces the three or four that actually matter.

Where plan-first pays most

Long outputs. Anything over about 500 words. The cost of a wrong outline is the whole draft.

Anything that will be run. A script, a spreadsheet formula, a bulk change. Read the plan before something touches real data.

Anything with a structure you will argue about. Reports, proposals, curricula. Structure is where most of the disagreement lives, and it is much cheaper to argue about six headings than about 2,000 words.

Multi-step work with tools. If a model is going to search, read files or call things, the plan tells you what it is about to do while it is still a sentence.

Where it is a waste

Short tasks. A 150-word email does not need an outline; asking for one doubles the interaction for no gain. Anything you will read in full and can judge instantly is faster to just produce.

Approving a plan is a real decision

Two failure modes to watch in yourself.

Rubber-stamping. A plan that looks reasonable usually is reasonable, so it becomes habit to say “yes, go”. Read the assumptions list, at least. That is where the surprise is.

Over-planning. Some people iterate on the outline four times and produce a beautiful plan for a document nobody needed. Two rounds is almost always enough; if the third round is still finding structural problems, the request itself is unclear and no plan will fix it.

The version for automated work

When a model has tools (running code, editing files, sending things) plan-first stops being a convenience and becomes a control.

Produce the plan as a numbered list of actions. For each, say what it changes and whether it is reversible. Wait for my approval. Do not act.

The reversibility column is the useful one. It sorts the actions into those you can skim and those you must read, and it is the shape of every sensible approval gate in automated systems. Anything irreversible (sending, deleting, paying, publishing) should have a human between the plan and the act, and that principle does not weaken as models improve.

The two-line habit

For anything substantial:

First: outline plus the assumptions that would change the result if wrong. Then stop.

Fifteen seconds to type, and it will save you an afternoon roughly once a fortnight.

A worked case

A consultant asked for a 3,000-word market overview and got something competent about the wrong market: the model had read “the education sector” as schools when she meant corporate training. She noticed on page two.

The plan version would have surfaced it in one line: Assumption: “education sector” means K-12 and higher education institutions. Ten seconds to read, five seconds to correct, and the twelve minutes of generation would have been spent on the right subject.

The interesting part is that her prompt was not badly written. It was ambiguous in a way that is invisible from the inside, because she knew what she meant. That is the normal case, not a lapse, and it is why the assumptions list beats a careful re-read of your own prompt.