Format and Examples — what good output looks like

The last two slots tell the model what its replies should look like. They do less work than Role and Constraints, but they are the difference between an assistant that "mostly behaves" and one that ships.

The Format slot

Format covers four things: length, structure, sign-off, and what to omit. Each one is a single line.

Format:
- 4 short paragraphs maximum, no bullet lists.
- No subject line, no markdown headings.
- No disclaimers or "as an AI" framing.
- End with "— Bayt Coffee team" on its own line.

Two rules of thumb for writing this slot:

1. Be specific about length. "Keep it short" is meaningless. "4 short paragraphs maximum" is enforceable.
2. Tell the model what to omit. The model defaults to adding subject lines, disclaimers, "let me know if you have any questions" sign-offs. If you do not want them, say so explicitly. Negative format rules ("no markdown headings") are often more useful than positive ones.

The Examples slot

Examples are optional. They are also the single highest-leverage slot when the output is hard to describe in words. One worked example is worth ten constraint lines.

For Bayt Coffee, Hagar adds one example to lock in the shape:

Example:
User: "My order is 2 days late."
Assistant: "Your order is 2 days behind schedule — that's not okay.

I'm pulling the tracking now and will email you a status update with a new delivery window by 6pm today.

If it doesn't move within 24 hours, I'll refund the shipping and add a free 250g bag to your next order.

— Bayt Coffee team"

Three things this example does that no constraint line could do as well:

• Shows tone in action. "That's not okay" is the warm-but-direct register the constraint slot only described.
• Shows structure. Three short paragraphs, no bullets, sign-off on its own line. Faster than reading the format rules.
• Shows the recovery pattern. Acknowledge → next step → fallback → sign-off. The model now has a template.

When to skip examples

Skip them when the task is genuinely simple ("classify this email") or when the constraints are already specific enough to lock in the output. Add them when the task involves voice, recovery patterns, or a multi-part structure that is awkward to describe in rule form.

A two-example prompt is usually enough. Beyond three, you are no longer writing a system prompt — you are writing a few-shot demonstration, and that belongs in the user message, not the system slot.

Next: the full Bayt Coffee system prompt, end-to-end, with the live model output. :::