Create effective prompts for business and technical tasks
Write prompts that give Claude everything a well-briefed colleague would need: audience, goal, format, constraints, source material, examples, and permission to say what it does not know.
Key points
- 1
Treat Claude like a brilliant new employee with no context on your organisation. The help centre's rule: write the prompt as if for someone with no background on the task, and if a colleague would be confused by it, Claude will be too.
- 2
Always state the audience, the purpose of the output, the format (email, table, bullet list, memo) and the length. "Write a launch email" produces a generic draft; "Write a 150-word launch email to existing customers announcing X, warm but not salesy, mentioning features A, B and C" produces a usable one.
- 3
Replace vague adjectives ("good", "professional", "thorough", "accurate") with measurable constraints: word limits, a fixed scale, a list of required facts, a named tone example.
- 4
Supply the source material. Claude cannot know your policy, job description, report or data unless it is in the chat, attached, or in a Project. Most "wrong answer" problems trace back to missing context.
- 5
Explain why a constraint matters. "Never use ellipses because the text will be read aloud by a screen reader" works better than "never use ellipses"; Claude generalises from the motivation.
- 6
Use examples (few-shot prompting) to steer format, tone and structure. Two or three relevant, varied examples of the finished output are one of the most reliable techniques; wrap them in
<example>tags so Claude can tell them from instructions. - 7
Give Claude a role when tone or perspective matters ("act as an experienced HR interviewer"), but remember a role is not a specification; it does not replace criteria, scales or formats.
- 8
For prompts that mix instructions, context, input data and examples, label the parts with XML-style tags such as
<instructions>,<context>,<input>, so nothing is misread as something else. - 9
Give Claude an out: tell it to say "the document does not cover this" or to flag uncertainty rather than estimate. This measurably reduces invented figures and citations. Restrict answers to the supplied source when accuracy matters ("use only figures that appear in the report, with the page number").
- 10
Ask for sequential, numbered steps when order or completeness matters, and ask for the output format explicitly (columns of a table, sections of a memo).
- 11
For long inputs (a large report or several documents), put the documents at the top and the question at the end, and ask Claude to quote the relevant passages before answering.
- 12
A prompt that works for its author often relies on private context. To share it with a team, externalise that context: show tone by example, replace judgment calls with numbers, and turn unverifiable rules ("nothing confidential") into checkable ones ("only names listed in the input").
- 13
Common distractors: adding "be accurate" or "MUST" instead of supplying facts; switching to a bigger model instead of fixing the prompt; asking Claude to rate its own confidence; retrying the same prompt unchanged.
Read the source
Test yourself on Create effective prompts for business and technical tasks
Ten questions, with the answer and explanation after each one.