Home / What a language model is, without the jargon
Meaning without a dictionary, and why order matters
Claude opens no dictionary, it associates each word with words of similar meaning and weighs each one against all the others at the moment of answering, which means the place of an instruction in the message changes its real weight in the answer.
Claude consults no dictionary to understand a word. It associates each word with the words of similar meaning it encountered during training, then, at the moment of answering, it weighs each word in the message against all the others to decide what matters most. This weighing is not uniform across the whole length of the message, the position of a piece of information changes the real weight it receives.
The beginning and the end are better retained than the middle
Anthropic has published a study on this exact point, carried out on very long messages. The central result fits in one sentence, placing an instruction right at the end of the message increases the chances that Claude remembers it at the moment of answering, because the end of the message carries more weight in the final calculation. The same study specifies that this finding is not a universal law, on another model tested, the observed relationship traced more of a dip shaped curve instead, with a beginning and an end equally well retained, and certain writing techniques, such as providing examples or asking for written reasoning before the answer, can actually work against a piece of information placed right at the end of a very long document.
Moving the instruction that really matters
The practical rule that follows from this stays simple despite the nuance, when a message goes beyond a few sentences, the most important instruction is best placed right at the beginning or right at the end, never buried in the middle of a long paragraph of context. Here is the same message, before and after this move.
Before: here is the minutes of the meeting, ten paragraphs of scattered notes,
summarise these minutes in three points for absent members, then the appendices.
After: summarise these minutes in three points for absent members.
Here are the minutes, ten paragraphs of scattered notes, then the appendices.
In the second version, the instruction opens the message, it does not wait to be found after ten paragraphs of context.
The principle seen in the previous lesson, a deeper model does not fix a poor structure, also applies here, the order of words acts on a lever distinct from the choice of model, within a single message.
The real weight of an instruction depending on its position in the message
A volunteer trainer writes a nine paragraph message for his assistant, he opens with the detailed minutes of the last meeting, continues with figures in an appendix, and finishes with the sentence summarise these minutes in three points for absent members.
Write in one sentence what this situation establishes, and in one sentence what it does not establish.
What this establishes: This message places the most important instruction in the very last position, the spot Anthropic's study identifies as one of the best retained by Claude.
What this does not establish: This does not establish that Claude actually summarised the minutes into three consistent points, since the situation stops before any answer.
The three most common miscalibrations
- Too broad This message guarantees that Claude retains every figure in the appendix as well as the summary sentence placed at the end.
- Too narrow Nothing can be affirmed about the recall of the final instruction from a single message written just once.
- Beside the point This message confirms that this trainer's assistant is configured on the fastest model in the range.
- Claude opens no dictionary, it weighs each word in the message against all the others at the moment of answering.
- The position of an instruction in a message changes its real weight, a piece of information placed right at the beginning or right at the end is better retained than one buried in the middle.
- Anthropic's study on this point specifies that the rule is not universal, another model tested shows a different curve depending on the writing technique used.
- Moving the most important instruction to the start or end of a message is a lever independent of the choice of model.
Take a long message you usually send to your assistant, identify the most important instruction it contains, and move it to become the very first or very last sentence before sending it again.
Every datable claim in this lesson links here to the public text behind it. A source that does not open proves nothing.
- Prompt engineering for Claude's long context window consultée le 2026-09-02