Make it commit to a plan you can reject

Reject the plan, not the document

A wrong 1,500-word document costs you fifteen minutes to read and a decision about whether to salvage it. A wrong six-line plan costs you thirty seconds and one sentence of correction.

So ask for the plan.

Before writing anything, give me: the audience, the single decision this document asks the reader to make, the section headings in order, and one line on what evidence goes in each section. No prose yet.

Read it. Fix it. Then:

Good, but the audience is internal, not the client, and section three should come before section two. Now write section one only.

Why it improves the draft, not just your review

There are two separate benefits and they are worth separating.

The first is the obvious one: you catch a wrong approach cheaply.

The second is mechanical. Once the plan exists in the conversation, the draft is generated conditioned on it. Every subsequent sentence is produced with the structure already present in the window, which pulls the output towards it. This is the same reason that asking for evidence before a verdict works: text produced earlier constrains text produced later, because there is only one process and it runs forwards.

A plan you approved is therefore doing more work than a plan you merely read.

What the plan exposes

The plan surfaces the assumptions the model would otherwise bury in fluent prose. Almost every time, at least one of these is wrong and you would not have noticed it in a draft:

• Who it thinks the reader is.
• Whether it thinks you are proposing, informing, or asking permission.
• What it thinks the strongest argument is.
• Which of your facts it intends to lead with.

The last one is the sharpest. If it plans to open with the cost saving and your actual case is about risk, the whole document was going to be pointed in the wrong direction, and you would have felt the wrongness only after reading all of it.

The honest limitation: a stated plan is not an explanation

Here is where a lot of confident advice goes wrong, so be careful with this.

The plan the model writes is text, generated the same way as everything else. It is a commitment device and a proposal you can correct. It is not a window into the computation.

There is a body of work on this — models given a subtle hint towards a particular answer will often take the hint and then produce reasoning that never mentions it, describing instead a chain of thought that sounds principled. The written reasoning and the factors actually driving the output can come apart, and you cannot tell from reading it.

What follows practically: use the plan to steer and to catch structural errors. Do not use it as evidence that the answer was reached soundly, and never present it to somebody else as an explanation of why the system produced what it produced. If a decision needs to be justifiable, the justification has to be yours.

When planning first is not worth it

• Short tasks. A three-sentence reply does not need an architecture.
• Pure style work. Rewriting a paragraph in plainer language has no plan; it has a paragraph.
• When you already have the structure. If you know the five headings, skip the plan and give them. You were the plan.

The overhead is one extra turn, so the test is simply whether a wrong structure would cost you more than that turn.

The same move, applied elsewhere

Plan-first generalises past documents, and the pattern is worth carrying:

• Before an analysis: "What steps would you take, in order, and what would each one need from me?" You often discover it needs a column you have not got.
• Before a formula: "Describe in words what the formula must do, then write it." The description is where you notice it has misunderstood which column holds the date.
• Before a long extraction: "Show me the output format for the first two rows only." Fix the shape before generating 400 of them.
• Before a meeting: "List what this agenda assumes has already been agreed." Frequently the most useful ten seconds of the week.

Each is the same move: make it commit cheaply, in a form you can reject, before it commits expensively in a form you have to read.