Minto Pyramid and Autoregressivity

How a person actually writes a technical text

Watch a patent attorney drafting a client letter about an office action. She writes the first paragraph, realises in paragraph four that the framing from paragraph two does not hold, jumps back, strikes a sentence, continues differently. In paragraph five she notices that the recommendation she had planned to put at the end is actually the anchor of the whole letter. So she restructures.

The back-jump is not a mistake. It is the craft. No experienced technical writer produces a complex piece in one sequential pass and leaves it untouched. Real writing practice is two to five loops through the material. Each loop makes the text clearer, because the writer now knows what the text says at the end.

For patent practice, this revision is not a side point. A response to an examination report needs to carry a clear answer: which claims are defended, which are amended, which are dropped. This answer sits at the end of the thinking process, not at the beginning. Anyone who wants to put it at the top of the letter must run the letter a second time.

The Minto pyramid as a top-down structure

Barbara Minto, working at McKinsey in the 1960s, observed that consulting memos were consistently read poorly whenever the recommendation sat at the end. Her answer, worked out in "The Pyramid Principle", is structural: the core statement sits at the top, three to five supporting pillars follow, and the details sit at the bottom. Readers who stop after the first paragraph still get the most important part.

This is more than a formatting trick. The pyramid requires the writer to formulate the core statement precisely before writing. In many cases, she can only do that after the fact, while reading her own draft. The pyramid therefore necessarily emerges on the second pass. Anyone who tries to write Minto-compliant in a single go either has to know the result up front or accept a text that corrects itself midway and stumbles because of it.

For a client letter responding to an office action this means: page one carries the answer and the three key points. A client who reads only that page knows how the matter will proceed. Anyone who reads deeper gets the evidence. This is the opposite of the classical opinion style with fifteen pages of build-up and the conclusion in the last paragraph.

What autoregressivity means for language models

A language model generates strictly from left to right. Formally, every token is a probability distribution conditioned on every token already written: P(token_n | token_0 … token_(n−1)). Once a token is placed, it stays. Within a single generation run there is no mechanism to go back to an earlier token and overwrite it.

Three hard consequences follow for longer texts.

1. No restructuring within a single answer. If the model notices at token 400 that the framing from token 20 is off, it has to build the rest of the answer around that mistake, or try to explain it away later. What a human does with three clicks in an editor (cut a paragraph, paste it on top) is impossible in a single model run.
2. The core statement lands at the end. The model often discovers it only while writing, because for an LLM, writing and thinking are the same operation. That is why long LLM answers tend towards a bottom-up structure: long build-up, conclusion at the end. Minto-compliant top-down texts are not reachable in one go unless the core statement was already explicit in the prompt.
3. Framings become prisons. If the answer starts with "The problem can be summarised in three points", the rest of the text must follow that frame, even when the model notices midway that four points or a different structure would be more precise. A human would strike the first sentence and start again. The model cannot.

What thinking models solve, and what they do not

Newer models such as GPT-5 with reasoning, Claude Sonnet 4.6 with extended thinking, or Gemini with thinking mode run an internal scratchpad pass before producing the visible answer. They plan, reject approaches, and reframe. The user does not see this, or sees it only in a collapsible panel. This measurably improves the first draft, especially for tasks that require planning (maths, code, multi-step argument).

What thinking models do not solve: once the visible answer begins, it is autoregressive again. Token by token, no going back. The thinking phase defuses the planning problem before writing. It does not defuse the revision problem after writing. The difference is structural.

Iteration as the structural compensation

If the model cannot restructure, someone else has to. The human reads the locked output, realises the recommendation should sit at the top, and sends a second prompt. "Rewrite with the recommendation first, then the supporting analysis, no preamble." The model writes a second time. This run can go top-down, because the answer is now known.

This is not "a feedback loop because the user changed their mind". It is the structural compensation for a capability the model lacks. Without that second pass, the output misses precisely the quality level that a human would call the "second draft".

Iteration runs on two levels. Macro-iteration happens between human and model: read output, new prompt, new output. Micro-iteration builds the same cycle programmatically, for example in drafter/reviewer setups or in agent architectures that generate several internal passes before the final answer. Both levels solve the same problem: they create a second run on top of the first.

Role shift: editor instead of author

In classical writing, one person is author and editor in a single head. When working with a language model, that splits. The author sits inside the model. The human role becomes editor and director. This is not less work than classical writing; it is different work. The leverage lies in editing: spotting the core statement, restructuring, cutting.

This role shift explains why practitioners who are good at editorial thinking produce noticeably better results with LLMs than those who try to craft a perfect one-shot prompt. The quality jump sits in the second, third, fourth round. Skipping that part gives away most of the tool.

What this means concretely for patent work

Three patterns follow from this diagnosis and are directly applicable in firm practice:

Cross-references: drafter/reviewer in separate chats is pattern 6 in Context Engineering. The duty of evidence in both sides' output belongs to Duty of Evidence. The prompt patterns for the individual roles are covered in Prompt Engineering Patterns.

Ready to use: the prompt tool Minto Pyramid Rewrite turns these principles into three templates (direct rewrite, skeleton first, structure extraction for long texts) — paste the original, set the context, copy the prompt.

Sources

1. Minto, Barbara: The Pyramid Principle: Logic in Writing and Thinking, 3rd edition, Pearson Education, 2009 (first edition 1978). Chapters 1 to 3 for SCQA and the pyramid construction.
2. Vaswani, A. et al.: "Attention Is All You Need", NeurIPS 2017. Foundational paper on the Transformer as the autoregressive architecture behind current language models. arxiv.org/abs/1706.03762
3. Bommasani, R. et al.: "On the Opportunities and Risks of Foundation Models", Stanford CRFM 2021. Sections on the structural limits of autoregressive generation. arxiv.org/abs/2108.07258
4. OpenAI: "Learning to Reason with LLMs", 2024. Technical background on reasoning models and internal scratchpads. openai.com
5. Yao, S. et al.: "ReAct: Synergizing Reasoning and Acting in Language Models", ICLR 2023. Prototype for iterative micro-loops in agent architectures. arxiv.org/abs/2210.03629