AI Guides › Workbench
By Nigel Guy · 7 min read
Most people write a long prompt the way they would brief a colleague: say what you want first, then paste the material underneath. It reads naturally, it works fine on a one-page email, and it quietly gets worse as the pasted material grows. The model reads your instruction, then wades through thousands of words with no question freshly in view, and the answer comes back vague or drifts to the wrong section.
The rule: put the long material at the top, the question at the bottom, and ask for quotes before conclusions.
This is the whole trick. It is free, it takes thirty seconds, and it is a placement habit, not a magic phrase. Below is the evidence for it, three layouts you can paste straight in, and the places it will not help.
Anthropic's own prompting guidance for Claude, in the section on long-context prompts, says to place long documents and inputs near the top of the prompt, above the query, instructions and examples. It describes this as improving performance across models, and adds that putting the query at the end can improve response quality by up to 30 percent in tests, especially with complex, multi-document inputs. The guidance frames this for large inputs, around 20,000 tokens and up (very roughly 15,000 words), so a two-paragraph paste will not show much difference.
Read that 30 percent carefully. It is Anthropic's figure from its own tests, it says "up to", and it applies to complex multi-document cases. Treat it as a reason to try the layout, not a promise about your files.
The same guidance gives two companion habits: wrap each document in labelled tags with its source, and ask the model to quote the relevant parts first, before it does the task, so it filters out the noise in the rest of the text.
OpenAI's GPT-4.1 prompting guide says something related but not identical: for long context, put your instructions at both the beginning and the end of the provided material, which they found performed better than only above or below. If you use the instruction only once, they say placing it above the context beats placing it below. So the vendors differ slightly. Where you are unsure, the safe compromise is the one in layout 2 below: material first, then the task, with a one-line pointer to the task at the top.
| Layout | What it does | Cost | Best for | Catch |
|---|---|---|---|---|
| 1. Single document, question last | Document at top, task and question below | Free; uses your normal plan allowance | One long report, transcript or contract | Easy to forget the question, so write it first and paste it last |
| 2. Tagged multi-document pack | Each source wrapped in tags with a name, task at the bottom | Free; uses your normal plan allowance | Comparing two to five files | Longer to set up; tags must be closed properly |
| 3. Quote-first pass | Model pulls verbatim quotes, then answers from them | Free; uses slightly more output | Anything you will rely on or check | Adds a step; quotes still need spot-checking |
There is no price to convert: the technique costs nothing beyond whatever your plan or API usage already charges. Check the current price of your own plan on the vendor's page.
Paste the document first, with a plain label, then the task. Fill in the square brackets.
<document>
[PASTE_YOUR_DOCUMENT_HERE]
</document>
You are a careful analyst reading the document above. It is a [DOCUMENT_TYPE, e.g. board report or interview transcript].
My goal: [WHAT_YOU_NEED, e.g. the three biggest risks the author understates].
A good answer: [FORMAT, e.g. a numbered list of up to five items, each with one sentence of explanation and the section it came from].
Rules:
- Use only what is in the document. If the document does not say, write "not stated".
- If my goal above is unclear or a detail is missing, ask me one question before answering.
- Before you reply, check each item against the document and delete any you cannot point to.
Fill in: the document, its type, your goal and the output format. Note the question sits after the material, which is the point.
When you have several sources, give each a name so the model can cite it and you can check it.
<documents>
<document index="1">
<source>[FILE_NAME_OR_TITLE_1]</source>
<document_content>
[PASTE_DOCUMENT_1]
</document_content>
</document>
<document index="2">
<source>[FILE_NAME_OR_TITLE_2]</source>
<document_content>
[PASTE_DOCUMENT_2]
</document_content>
</document>
</documents>
Role: you are a [ROLE, e.g. procurement reviewer] comparing the documents above for [AUDIENCE, e.g. a finance director].
Task: [QUESTION, e.g. where do the two quotes disagree on price, scope or delivery dates?]
Output: a table with the columns Point, Document 1, Document 2 and Source. Name the document by its source tag in every cell.
Constraints: no outside knowledge; flag any figure that appears in only one document; if a document is missing something needed to answer, say what is missing rather than guess.
Self-check: before answering, confirm every row cites a document and that nothing in the table is invented.
Fill in: titles, the pasted text, role, audience, question. This structure is adapted from the tag format in Anthropic's long-context guidance.
Use this when the answer will drive a decision, because you can check the quotes against the original in seconds.
<document>
[PASTE_YOUR_DOCUMENT_HERE]
</document>
Step 1: Find the passages in the document most relevant to this question: [YOUR_QUESTION]. Copy up to [NUMBER, e.g. 8] of them word for word inside <quotes> tags, numbered. If nothing is relevant, write "no relevant passages" and stop.
Step 2: Using only those quotes, answer the question in [FORMAT, e.g. under 200 words]. Refer to quotes by number.
Step 3: List anything the question needs that the quotes do not cover.
Do not paraphrase inside the quotes. If you are not sure a passage is exact, leave it out. Ask me if the question is ambiguous.
Fill in: the document, the question, the number of quotes and the answer format.
Take one real long document and one question whose answer you already know. Run it twice in fresh conversations: question first, then question last. Compare against what you know is correct, and note which answer cites the right section. One test proves little because outputs vary run to run, so repeat it on a second document before you change your habits.