Claude — prompt engineering

Be specific and direct

Assumptions fill the gaps. Close the gaps yourself.

• State what, for whom and in which format — not just the topic.
• Positive instruction beats negative: write "answer in prose" rather than "no bullets".
• Give context on why — the purpose drives tone and depth.
• Set explicit limits: length, reading level, what may be left out.

# weak
Write about zero trust.

# strong
Write a 200-word paragraph on zero trust for a
non-technical management team. Focus: why the
perimeter model fails. Prose, no jargon.

XML tags to separate

Claude is trained to respect tags. Use them to split instruction from data.

• Wrap input, examples and rules in named tags.
• Refer to tags later in the prompt by name.
• Consistent tag names make long prompts robust.

<instruction>
Summarise the report in 3 bullet points.
</instruction>

<report>
{{paste text}}
</report>

<format>
Each bullet max 15 words.
</format>

Few-shot & multishot

The fastest route to consistent format and tone.

• 2–5 examples cover most edge cases.
• Show both input and desired output — not just the output.
• Include a hard or atypical example, not only the easy one.
• Keep examples consistent in style — deviations get copied.

Chain of thought

Let the model think before it concludes on complex tasks.

• Ask explicitly to "think step by step" for analysis, maths, logic.
• Structure thinking in a <thinking> tag, answer in an <answer> tag.
• Do not use it for simple lookups — it costs tokens with no gain.
• Want to hide the reasoning from the end user? Ask for it separately.

Think the problem through in <thinking>.
Give only the final recommendation in <answer>.

System prompt & persona

Set the role in the system prompt — it hits expertise and tone in one move.

• System prompt = who Claude is; user turn = the task.
• A concrete role beats a vague one: "DFIR analyst reviewing an IR report".
• Put durable rules (tone, language, output) in the system, not in every message.

// system
You are an experienced governance consultant.
Answer concisely, in English, and always with
concrete actions rather than general advice.

Steer output directly

Prefill and format control remove leading filler.

• Prefill the assistant reply with e.g. { to force clean JSON.
• Prefill with a tag to skip the preamble.
• Ask for an exact schema when output is parsed by machine.
• "Answer only with …" when you want to avoid helpful chatter around it.

# force JSON, no chatter
Answer ONLY with valid JSON in this schema:
{ "risk": "high|medium|low",
  "rationale": "string" }

When the context is large

Order and placement matter when you feed whole reports in.

Common mistakes