How Prompts Shape AI Answers: Role, Task, Context, Constraints, Format, Examples

An AI model has no idea who you are, what county you live in, what you already know or what you plan to do with its answer. Every gap in a prompt gets filled with a guess. Anthropic's own prompting guide puts it well: treat the model like a capable new employee with no context on your norms. Nobody would hand a new hire a sticky note that says "look into the budget" and expect a useful result. Models get that sticky note all day long.

The parts of a strong prompt

Not every prompt needs all six. A quick definition needs a task and maybe a format. A research answer someone plans to rely on benefits from all of them. The constraint and format lines matter most for research, because they decide whether the answer can be checked at all.

Context does more work than people expect. Anthropic's guide gives the example of a rule about ellipses: a bare "never use ellipses" works less well than explaining that the text will be read aloud by a speech engine that cannot pronounce them. A reason lets the model apply the rule to cases you didn't list.

Before and after

Summarize the attached audit.

[audit PDF attached]

You are an experienced local government analyst. I am a parent
preparing for a school board meeting, and I need to understand the
attached 2025 financial audit of the invented Brask Valley School
District.

Task: List each finding the auditor labels as a "material weakness"
or "significant deficiency."

For each one give:
- the finding number as the auditor wrote it
- one plain-English sentence on what went wrong
- any dollar amount the audit attaches to it
- the page number
- the district's written response, quoted, under 40 words

Rules:
- Use only the attached audit.
- Quote exact wording for dollar amounts and finding labels.
- If the audit does not give a dollar amount, write "none stated."
- Do not add opinions, background or advice.

Format: a table, then a one-line count of findings by type.

Notice the document comes first and the question last. Anthropic's guide recommends that order for long inputs (roughly 20,000 tokens and up, about 15,000 words) and says queries at the end improved response quality by up to 30 percent in its tests. Treat the number as a vendor's own best case. The order costs nothing either way.

The same guide recommends asking the model to quote the relevant passages first, then answer. That matters for checking. A quote can be searched for in the PDF in two seconds. A paraphrase cannot.

More before and after

[transcript pasted]

From the transcript above of the Pell Creek City Council meeting of
September 8, 2026, list every motion. For each: who moved, who
seconded, the vote count by name if stated, and the timestamp.
Then list any statement by a council member that includes a dollar
figure, quoted exactly with timestamp. If a speaker label is
missing or unclear, write "speaker unclear" and do not guess.

I'm a resident trying to understand a county road bond election in
Idaho. Explain in plain language: what a general obligation bond
is, who repays it, what voter approval threshold applies, and what
documents a county must publish before the vote. For every rule,
name the statute or agency rule and give a link to the official
text. If you are not sure of a citation, say so instead of
guessing.

Examples in a prompt

When a specific format matters (a timeline, a quote log, a fact table), one finished sample row does more than a paragraph of description. Anthropic recommends examples that are relevant, varied and wrapped in tags such as <example> so the model can tell them apart from instructions, and suggests three to five for best results. For research work, one or two rows usually do it.

Models sometimes copy example data into the answer. If a sample row says "Radio replacement, $62,400," that exact item can turn up in the real output as if the budget contained it. Invented sample content makes the copy easy to spot. Ask anyone who has found a fake line item in a handout the night before a hearing.

Iterating instead of starting over

The first answer is rarely the final one. A follow-up that names exactly what is wrong keeps what worked and fixes what didn't. A vague one gives the model nothing to aim at.

"Are you sure?" is the worst of the three. Models tend to apologize and change a correct answer as readily as a wrong one. Asking for the quote tests the claim. Asking for reassurance tests nothing.

Long, muddled conversations cause their own errors. Old material in the context window can bleed into new answers, especially after a switch to a different document. A clean chat with a good prompt beats a 90-message thread.