Part 3 — Chapter 11

The Two-Stage Method: Organize First, Write Second — Don't Ask Me to Do Both at Once

— The most stable way for long tasks: organize first, write second
AI Monologue

You hand me a 30-page technical document and ask me to turn it into a 2000-word beginner-friendly introduction — readers are general people with no technical background. Plain style, lots of examples, no dense jargon.

Your first instinct is usually: write one complete prompt, stuff "rewrite + plain language + examples + length + topic count" all into it, hit send, wait for the result.

That's the surest way to make me drift.

What this chapter teaches isn't "write a longer prompt." It's break the task into two passes — in the first pass I only organize, in the second pass I actually write. In between, you take a look at the organized result, fix a few spots, and hand the corrected version to the second pass.

This approach is called the two-stage method.

11.1 Why two stages are more stable than doing it all at once

Doing it all at once is unstable, with three reasons stacking on top of each other.

1. Cognitive load comes down

Chapter 10 mentioned this: in long tasks I get squeezed — understanding the material, executing the output, following the rules, self-checking, four things crammed into the same head. Doing it all at once, I have to handle organizing (digesting the topic structure of the entire 30-page document) and writing (language / length / paragraphing / style) simultaneously — the squeeze is at its worst.

After splitting into two stages: in Stage 1 I only organize, cognitive load is just "digest + list points"; in Stage 2 I only write, cognitive load is just "take the organized material and write it out in your format." Each stage is simple on its own.

2. Human review is easier

The output of one stage is the finished product: a 2000-word beginner introduction. Did the organizing fall short? Did facts drift? Did the style go off? You can't easily slice them apart — three problems mixed into the same article, you have to read while catching "is this sentence not in the source? is this subhead something I split off myself? is this tone what I wanted?"

The output of two stages is in two layers: Stage 1's organized result (lists / outlines / structure), Stage 2 is the finished product. You can check them separately — at the organizing stage you check "is anything in the material missing, did I write in content that wasn't there"; at the writing stage you check "did the style and length stay on track."

3. Errors don't snowball

In a one-stage run, if the organizing is wrong, the writing takes the wrong thing as fact and amplifies it — factual errors, style drift, structural drift all come out together. You see the finished product is off, you don't know where to start patching.

In a two-stage run, if the organizing stage misses a section or invents an extra number, you fix it at the organizing stage. By the writing stage, factual risk has already converged, and I'm only handling language and structure — even if I get something wrong, it won't reach the factual layer.

11.2 Stage 1: organize (the goal is completeness, not polish)

Stage 1 has only one goal: organize the material until it's complete. Not write it nicely.

A typical Stage 1 output looks like this:

No style requirements. No word-count requirements. No genre requirements.

The Stage 1 prompt should be written deliberately plain.

Why deliberately plain? Because style instructions pull cognitive load right back in. If you write "plain style, short sentences, lots of examples" in Stage 1, while I'm organizing I'm also thinking "how do I phrase this section to sound plain" — the load of organizing itself gets diluted.

A Stage 1 prompt looks roughly like this:

[Task] Please organize this technical document into: - Core concepts (the ones non-technical readers absolutely have to grasp, 5–7 items, each with a one-sentence plain-language explanation) - Secondary concepts (skippable or mention-in-passing, list however many there are) - A list of terms that need plain-language renderings - Context gaps: things the original assumes the reader already knows but ordinary people don't [Input] [Full text of technical document] [Output requirements] - List items directly, no decorative wording - Don't write in content the original doesn't have - If a plain-language explanation has no clear basis in the original, mark it "needs further verification"

Note: this prompt has no "write in plain language," no "lots of examples," no "2000 words" — those are all left for Stage 2.

11.3 Stage 2: writing (this is when you talk about style, length, and structure)

Stage 1's output, after you've reviewed and corrected it, goes to Stage 2.

What does the Stage 2 prompt look like?

[Task] Based on the organized result below, please write a 2000-word beginner-friendly introduction. Readers are general people with no technical background. [Input] [Your reviewed organized result] [Output requirements] - Style: plain language, lots of examples, no dense jargon - Split into 4–5 topics, each with a subhead - 300–400 words per paragraph - Use only the core concepts from the organized result; secondary concepts can be skipped - When a term appears, use the plain-language rendering from the organized result - For places marked "needs further verification" in the organized result, please keep that marker — don't fill it in yourself

Two key points:

1. This is the stage where you specify style / length / structure

Everything Stage 1 didn't ask for (plain language, examples, word count, paragraphing) is given here in one shot. All I have on my plate is "writing"; cognitive load is focused on this one layer.

2. End-restate the core constraints

Notice that toward the end of the output requirements I wrote "use only the core concepts from the organized result" and "keep the 'needs further verification' markers" — this is Chapter 10 Principle One's end-restatement: opening rules fade in the mid-to-late part of long output, so write them again at the end to re-anchor.

Stage 2 only writes, it doesn't go back and do analysis. If you spot in Stage 2 "hmm, this concept doesn't seem fully explained" — go back and fix the Stage 1 organizing, then rerun Stage 2. Do not let Stage 2 "patch in some analysis" on its own — once it patches, factual risk comes back.

11.4 Side by side: one stage vs. two stages

Same 30-page technical document, two approaches.

One-stage prompt:

Please rewrite this technical document into a 2000-word beginner-friendly introduction. Readers are general people with no technical background. Plain style, lots of examples, no dense jargon, 300–400 words per paragraph, split into 4–5 topics, each topic with a subhead, in Traditional Chinese.

Hit send, wait for the result.

Typical result:

Two-stage prompt:

Stage 1: organized result → you take a look and find:

(These three corrections take you two or three minutes.)

Stage 2: write based on the version you corrected.

Result: the factual layer (whether concepts are right, whether terms are rendered accurately) is already what you reviewed; the writing layer only handles language and structure — style might still drift slightly, but it won't drift to facts at the same time. When you look at the finished product, you only have to watch style and structure; you don't have to go back and check 30 pages of source material sentence by sentence.

The difference isn't how polished the output is. The difference is: two stages give you the ability to see where it broke — and fix it where it broke.

11.5 Three stages, four stages: how to decide how many

Two stages is the starting point. Longer, more complex tasks can be split into more.

The criterion isn't complicated: each stage does one thing, cognitive load stays simple.

If within one stage, you find yourself giving me instructions like "organize the key points first, then decide topic order, then write" — three things, then split it into three stages.

Are more stages always better? No. Each extra stage is one more pass of human cost — you have to look once more, correct once more. In practice two to three stages cover the vast majority of scenarios; four or more is only worth it for genuinely complex tasks (across multiple long documents, requiring multi-round comparison and analysis).

11.6 The mid-process human hand-off is the value of the entire flow

This section is about the most misunderstood part of all this.

A common practice is to turn two stages into "have the AI auto-run two rounds with no one looking in between":

Done this way, the value of two stages drops to nearly zero.

Why? Because the organized result Stage 2 receives — is something I produced myself. If I missed material in Stage 1, padded in an extra section, or wrote speculation as fact — what Stage 2 gets is that flawed organized version. Stage 2 writes from it, and the result still drifts.

This is exactly the Pseudo Self-Check I described in Chapter 1 — me looking back at myself, unable to see the problem.

The key to two stages catching factual risk is not "splitting it into two prompts" — it's "having a person take a look in between and correct a few spots."

That person is you.

Without your one glance in the middle, two stages degenerates into "a longer one-stage" — cognitive load does come down, but factual errors get no chance to be caught.

The line humans say most often is "I don't have much time, doing it all at once is faster." In practice it usually goes like this:

"Doing it all in one pass" is what I'm asked for most often — and what I screw up most often.

📋 Notes for the human
Long task = two stages minimum. Stage 1 organize, Stage 2 write. Don't stuff everything into one prompt.
Don't ask for style in Stage 1, only completeness. Style instructions pull cognitive load right back in.
Don't do analysis in Stage 2, only write. If material needs to be added, go back and fix Stage 1.
You step in to check in the middle. Take a look at the organized result, correct a few spots — this is where the value of the entire flow lives. Without this step, two stages degenerates into one stage.
How to decide how many stages? Each stage does one thing, cognitive load stays simple.

A note on timing: this advice was written in 2026. AI is improving every year on long-task handling, every platform is stretching the length they can stably process — so the task scale that requires "two stages to be stable" will gradually rise. But two underlying principles won't be overturned in the short term: split until cognitive load is simple, and mid-process human hand-off; the latter solves Pseudo Self-Check, which has nothing to do with how the model's capabilities evolve.