prompt-engineering

Prompt engineering

Prompts aren't magic incantations. They're specs for a system that happens to read English. Clarity, constraints, and structure matter more than clever phrasing.

The five-part frame

1. Task — what you want done, stated concretely.
2. Context — what the model needs to know that isn't in the task.
3. Input — the actual content to operate on (clearly demarcated).
4. Format — how you want the output structured.
5. Examples (optional but powerful) — zero, one, or few shots of input → output.

Most bad prompts skip one or more of these. Most good prompts are boring and explicit.

Task

Be specific. "Summarize this" is vague. "Summarize this support ticket in 3 bullets, each ≤20 words, focusing on the customer's core problem and what they've tried" is specific.

Rule: if a human contractor couldn't deliver from your task description, the model can't either.

Context

Everything the model needs that isn't in the task:

• The persona (who is the model being, for what audience).
• The domain (what jargon is okay, what assumptions it can make).
• Constraints the model doesn't know from task alone.
• What not to do (negative constraints are often as important as positive ones).

Don't pad with irrelevant context. Every extra token is noise, cost, and distraction.

Input

Wrap the content in clear delimiters so the model knows where task ends and content begins:

<ticket>
{customer ticket text}
</ticket>

XML-style tags work well with Claude and are unambiguous. Use them.

Format

Tell the model exactly what the output should look like:

• JSON with specific fields.
• A markdown table with specific columns.
• A numbered list with exactly N items.
• "Only the final answer, no preamble."

Models default to conversational output. If you want structured, say so.

Examples

For non-trivial tasks, examples beat instructions:

Example 1:
Input: "I can't log in, it says my password is wrong"
Output: {"category": "auth", "severity": "medium", "next_step": "reset_password_flow"}

Example 2:
Input: "The app crashes when I open the orders page"
Output: {"category": "bug", "severity": "high", "next_step": "escalate_engineering"}

Now classify:
{new input}