Absorb technical or regulatory documentation without losing the thread
A 180-page manual, a technical standard, a long contract. This course shows what to do when the document is big and the decision depends on one specific passage inside it.
- 01 · Master a new subject
- 02 · You are here
- 03 · Prepare a decision
- 04 · Cut task time
Read a long document in three steps, requiring every piece of information to come with the passage copied out and the page it sits on, and throw away anything you cannot find in the document.
Sign up once to unlock this course
Course 01 is open. For courses 02 to 04 we ask for six fields. It is a single signup: done once, valid for the whole Academy. It is not a free diagnosis and it does not trigger automatic sales contact.
Dropping in the whole document at once is the worst way to read it
The temptation is strong: you get a 180-page manual, a standard or a long contract, drop the file into the tool and ask for a summary. The answer comes back organised, with bullet points, and it looks like it covers everything.
It covers the beginning and the end well, and the middle badly. And what decides is almost always in the middle: the exception, the deadline, the condition that changes the meaning of everything before it.
What gets lost in the middle
A group of researchers measured this directly. They wrote questions whose answer sat in one specific document, hidden among other documents that did not matter. Then they moved where that document sat in the pile: first at the start, then in the middle, then at the end.
The result has the shape of a U. The tool gets it right more often when the information is at the start or the end of what it received, and gets it wrong far more when the information is in the middle. And the bigger the pile, the worse the result — including in the tools advertised precisely for handling a lot of text.
Pasting the whole document and asking for a summary puts the passage that decides in exactly the part where the tool gets it wrong most.TheNeil's reading of the result above
It does not mean the tool is no use for long documents. It means it is good at finding and rewriting when you tell it where to look, and bad at deciding on its own what mattered across two hundred pages.
Three steps, every piece of information with an address
Instead of one reading, three steps. Each one has its own question and ends in something different. And none of them accepts information without saying where it sits in the document.
| Pass | What you ask for | Required output |
|---|---|---|
| 1. Map The organisation, not the content | How the document is divided: which chapters exist, what each one covers, and where the definitions, deadlines, exceptions and penalties are. | A list of chapters, one line each. No conclusions yet. |
| 2. Slice One chapter at a time | One chapter only, and for each piece of information: the passage copied word for word and where it sits — page, article, clause. | Each piece of information with its passage beside it. If the tool cannot copy out the passage, the information does not go in. |
| 3. Check You, in the document | Nothing. This step is yours. | Search for the passage in the original file, with your own PDF reader. Whatever you cannot find is discarded, not corrected. |
Information without a copied passage that you can find in the document is a draft, not information. That holds even when it looks right — and especially when it looks right.
The quote that exists and does not say that
There are three ways this goes wrong. The second is the most dangerous, because it looks as if it has already been checked.
| Failure mode | How it presents | How to detect it |
|---|---|---|
| The passage does not exist | It arrives with correct formatting and a clause number that looks real, but it is not in the document. | The search finds nothing. This is the easiest error to catch. |
| The passage exists, the conclusion is swapped | The passage really is there, but the conclusion attached to it came from another chapter, or ignores the exception that appears right after. | Read the whole paragraph around the passage, not just the quoted line. |
| Something is missing, and nobody said so | Everything the summary says is right. It just never mentions the condition that inverts the conclusion. | Ask on purpose which exceptions, deadlines and conditions exist in that chapter, and compare against the map from step 1. |
The clause on page 47
A team has to say whether a new way of serving customers meets a requirement in the contract. The contract runs to 90 pages. The tool summarises the obligations and the team concludes everything is fine.
What did not appear in the summary was a condition, in the middle of the document, limiting the obligation to one specific service channel. Given the summary, the team's conclusion was right. Given the contract, it was wrong.
The middle of the document is where information gets lost
Where a passage sits inside what you hand over changes how much it gets used. That is what defines the size of the slice in step 2.
Where the information sat, and how much it was used
The horizontal axis is where the relevant passage sat inside what was handed to the tool. The vertical axis is how much the answer relied on it. The U shape is the finding: the start and the end get used well, the middle sinks.
Why is handing over the whole document the worst way to read it?
Which of the three ways of going wrong is the most dangerous?
What can a single reading deliver?
Try to answer from memory before you click. Same as course 01: trying to remember makes it stick, rereading only raises your confidence.
Where the argument comes from
- Nelson F. Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni and Percy Liang, “Lost in the Middle: How Language Models Use Long Contexts”, Transactions of the Association for Computational Linguistics, vol. 12, 2024, pp. 157–173. How the test was run, according to the paper: two kinds of task where the tool has to find the right information inside the text it received — answering questions using several documents, and locating a pair of items in a list — while moving where the relevant information sat. What the paper found: the tool gets it right more often when the information is at the start or the end, and gets it wrong far more when it has to reach into the middle, including in tools built for long texts; the result also gets worse as the amount of text grows. Verified 6 Aug 2026: authorship, journal, volume, pages, DOI and abstract wording confirmed on the ACL Anthology, MIT Press and arXiv (2307.03172).
The next course
You know how to study a subject and read the source. Course 03 uses that for the situation where study has to become a position: the meeting where you defend a recommendation.