Instruction order and placement: where things go in a prompt

What instruction placement is

Instruction placement is the practice of deciding where each part of a prompt goes: role and priorities, reference material, examples, and the actual question. The same words in a different order can produce different answers, because language models read strictly left to right and weigh positions unevenly. Two forces are at work, and the research on placement is mostly about which one wins in a given setup:

• Primacy: what comes first frames everything after it. In the instruction-overload study, models given many simultaneous requirements did not mix them up, they skipped them, and the skipped ones came from the tail of the list.
• Recency and causal reading: what comes last sits next to the point where the answer is generated, and only later tokens can "see" earlier ones. In the positional-failure study, moving the same question from the end of a long context to its middle dropped one model from 96% to 8%.

A workable rule reconciles both: put stable priorities (role, main goal, must-have format) at the top, put material in the middle, and put the specific task, the current data and the answer options at the end. When the context is long, say the task twice.

Why order changes the answer

Decoder models apply a causal mask: each token attends only to what precedes it. The CQO vs QOC study shows what that means in practice. With context, then question, then options, the options are encoded after the model has read the evidence. With question, options, then context, the options are encoded blind. Across 21 decoder-only models on four multiple-choice benchmarks, context-first was 14.7% more accurate on average, and accuracy in the options-first format matched having no context at all. The context's share of the prediction was 79.7% in one order and 33.5% in the other.

The end of the prompt matters for a second reason: proximity to generation. In the positional-failure study, 76% of errors with a mid-context task were answers copied from similar neighbouring examples, not reasoning failures; at the end position such matches were 22%. The model did not lose the ability to solve the task, it lost track of which task was being asked.

Primacy explains the other half. The primacy-bias analysis of 39 models argues that early tokens accumulate attention from every later position, so a requirement stated first keeps pulling even after you revise it in the chat. That is why a corrected value should be restated right before the question, not left "above".

Placement patterns that tested well

1. Task last. Put the document first and the instruction after it. The Q-Last study found that when the question opens a multi-turn case, models answer early: more than 55% of answers came within the first two messages. Held back until the question arrived at the end, models fixed wrong hypotheses 10 times more often than they broke right ones. An explicit "don't answer yet" barely helped; structure did.
2. Context before options. For any choice task, list criteria and facts first and the options last. If a fixed format forces options first, repeat them after the context (QOCO), which recovered 8.2% in the CQO study.
3. Repeat the task or the whole prompt. The Prompt Repetition study sent the query twice and recorded 47 wins out of 70 tests and 0 losses on non-reasoning models; on a find-the-name-in-a-list task, accuracy rose from 21% to 97% with triple repetition. In the positional-failure study a copy of the task at the end restored accuracy to within ±4 points of the ideal for all 9 models at 8K tokens. Related forms: asking the model to restate the question before reasoning (+10.4 points for DeepSeek-R1), asking the question before and after an image, and placing retrieved documents twice so facts are read before reasoning colours them.
4. Examples first, or next to the code. The position-effect study reports more stable few-shot answers when examples precede the question. For code, the co-located tests study found doctest examples inside the docstring linked to the function 2.8-4.4 times more strongly than tests at the bottom, with accuracy ranging from 0% to 100% depending only on placement.
5. Rank instructions by importance. Since the tail of a long rule list is what gets dropped, order constraints from critical to nice-to-have and keep the list short.

Primacy vs "question last": when each holds

The apparent contradiction. One line of work says "most important first"; another says "question last". They measure different things. Primacy studies look at compliance with many simultaneous constraints and at which version of a fact the model keeps. Recency studies look at whether the model finds and answers the right task among a lot of material. The long-context anatomy study combines them in a sandwich: instruction before the document, key requirements repeated after it.

The details still conflict. The lost-in-the-later study reports that models process the end of a long context worse and recommends instructions first, context after, plus an explicit "cover all parts evenly". The VeriCode study found that instructions in the middle of a list were followed 5-10% worse than those at the start or end, and that adding five style rules cut functional correctness by 6%. So the middle is the weak spot in almost every account; the ends disagree by task.

Where placement matters less:

• Some models barely care. In the positional-failure study DeepSeek-R showed no drop; in the long-context anatomy study Gemini and Deepseek held steady at every position while ChatGPT-5-mini dipped from 100% to 80% at the halfway point.
• Reasoning modes. Prompt repetition gave minimal gains when models already reason step by step.
• Short prompts. Most effects grow with context length and are small on a few paragraphs.
• Model updates. In the co-located tests study Code-LLaMA preferred separate tests, and one model family lost and regained the behaviour across versions. Re-test your order after switching models.

{context_part_1}

{intermediate_instruction}

{context_part_2}

----------------------------------------

Reminder, your task is: {question_or_instruction}

I'll give you the information in parts. Please don't draw a final conclusion until you've seen all the data; I'll ask my question at the end.

[Block 1: {general context / who's involved / what the situation is}]

[Block 2: {details of the situation, history, conditions}]

[Block 3: {specifics, nuances, constraints}]

[Block 4 - key data: {numbers, concrete facts, measurable indicators}]

Now the question: {your main question}

Question: {question}
A. {option_A}
B. {option_B}
C. {option_C}
D. {option_D}

{context}

Choose from: A | B | C | D

Answer: