The Ladder Rule

The problem with all-at-once requests

The bigger the task, the more likely the AI gets lost along the way. It starts strong, but by the third paragraph it's already forgotten what you asked for in the first.

This happens because the AI generates the entire answer in one pass. If the task has many parts, it dilutes its attention across them — and hands back something that reads like a finished document but is average everywhere.

There is a second, less obvious effect: when everything arrives at once, you cannot tell where the mistake is. Fixing turns into redoing.

The Ladder Rule (in 30 seconds)

Think of your task as a ladder. Instead of jumping from the floor to the top, you climb one rung at a time. Each rung is a separate prompt, and each prompt builds on what the previous one produced.

Don't do this: "Create a complete content strategy for the next quarter."

Do this:

1. First: "List the 3 main themes we should cover this quarter."
2. Then: "For theme 1, suggest 5 article titles."
3. Then: "Develop article 1 with an intro, 3 key points, and a conclusion."

Every rung has a success test: the theme list comes back clean, the five titles hold up, the piece has the three requested parts. Without that test, you cannot tell when to stop climbing.

Real example: product launch plan

Four prompts, each one building on the previous. In the end, you have a complete plan — and every part got the AI's full attention. If rung 3 comes back weak, you redo only rung 3. Rung 1 still stands.

How to size each rung

The right size depends on how well you already know the task. Two poles keep you calibrated:

• If you already know what the answer should look like, a rung can be longer: you can judge the output at a glance.
• If it is your first attempt, keep rungs short: less text to unpick when the direction is wrong.

The official prompt engineering guides from OpenAI and Anthropic describe the same principle under different names: break the task into subtasks, one at a time, and feed the output of one into the input of the next. The ladder is that idea with a name you can remember.

The rule of thumb is short: if the whole output of one prompt has to be pasted into the next, that is a rung. If you have to paste two different things, you sized it wrong.

The golden rule

If a prompt doesn't fit in three sentences, split it in two. Two precise prompts beat one confused one every time.

With practice, you'll learn the right size: not so small it becomes micromanagement, not so large the AI gets lost. And when the work is writing, the perfect prompt formula works as a template for each rung.

When not to climb

A factual question needs no ladder. Convert a currency, summarise a ten-line text, define a term, format a table: one prompt is enough, and extra rungs only delay the result.

The ladder pays off when the work has order of dependence — plans, reports, research, scripts, proposals. There, time spent on short prompts comes back on the first round of corrections you don't have to redo.

Frequently asked questions

How many rungs should a task have?

There is no fixed number. Start with three: definition, production, review. If a rung comes back too generic, split it; if the first one lands, move faster after it.

Do I have to start a new chat for every rung?

No — the opposite. Keep the same conversation, because each prompt leans on the one before it. Open a new chat only when you change tasks or when a long history starts pulling the model off course.

What if a rung in the middle goes wrong?

Fix it there and move on: with a success test per step, the mistake shows up the moment you read the output, not in the final document. If it repeats, rewrite that rung with tighter constraints.

Does the Ladder Rule apply when a team shares one AI thread?

It does, and it helps more. When several people read the same sequence of prompts, they understand the process without a walkthrough. To build those sequences together, see the piece on generic answers and how to respond to them.

The whole method fits on one line: one prompt, one test, one rung. Do that for a week and the right size stops being theory — it becomes reflex.