Prompt Engineering with Claude

The parts of a good prompt

A reliable prompt has five parts. The role or task states what Claude is doing. The context gives the background it needs. The input is the real material to work from. The constraints state the rules: length, tone, format, and what must not be invented. The output format states the structure of the answer. Missing any one part forces a guess, and guesses are where inconsistent results come from.

Write the goal first, before the details. If you cannot state the goal in one sentence, the prompt is trying to do two things at once. Split it. Two focused prompts beat one prompt that asks for a summary, a rewrite, and a plan in the same breath.

Be specific about constraints

Constraints are where prompt quality is won or lost. State the length, the audience, the tone, and the format. Say what to do when information is missing: mark it as to confirm, leave it out, or ask a question, rather than inventing it. Say which facts must appear and which claims must not be made. Each constraint removes a decision Claude would otherwise make on its own, and each removal makes the output more predictable.

Constraints also include what not to do. Do not use hype language. Do not add a conclusion if none was asked for. Do not repeat the input back. Negative instructions are useful, but keep them short: a prompt with twenty prohibitions is hard for anyone, human or model, to follow.

Use examples and structure

One example of the output you want is worth a paragraph of description. Paste a short example and name what to copy from it: the structure, the tone, or the level of detail. For repeated work, ask for a fixed structure every time, such as headings in a set order, a table with named columns, or a summary followed by action items. Fixed structure makes results easy to compare, easy to check, and easy to paste into your own documents.

For longer tasks, break the work into steps inside the prompt: first extract the key points, then organise them, then write the draft. Step-by-step prompts produce more reliable results than asking for a finished piece in one leap, because each step can be checked.

Test, then refine

A prompt is not finished when it works once. Run it on three or four different inputs, including a difficult one, and note where the output varies or goes wrong. The fix is usually a missing constraint or an unclear instruction, not a new trick. Add the rule that would have prevented the failure, test again, and keep the version that works. Save that version as a template with placeholders for the parts that change, so the next use starts from a proven prompt instead of a blank page.

Keep a small library of your best prompts, each with a note about what it is for and one example of good output. Over months, that library becomes more valuable than any single prompt, because it captures how you actually work.

Common prompt mistakes

The most common mistakes are easy to spot once you know them. Asking a vague question and expecting a specific answer. Giving no material and expecting accurate facts. Asking for everything at once. Accepting the first answer without checking it. And copying prompts from strangers without testing them on your own material. Your prompts should be built from your work, tested on your inputs, and judged by your results. That is the whole discipline, and it is enough.

Summarise this document for a busy manager.
Audience: a manager who has 2 minutes to read it.
Output structure:
1. Decision needed (one sentence, or "None" if no decision is needed)
2. Key points (no more than 5 bullets)
3. Action items (each with owner and date, or "To confirm" if missing)
Rules:
- Use only information from the document below. Do not add outside facts.
- If a date or owner is missing, write "To confirm" instead of guessing.
- Keep the whole summary under 250 words.
- Plain language, no hype.

Document:
[Paste the document here]

Frequently asked questions