Tone, length and structure — and avoiding unnecessary instructions

Tone: describe it in observable terms

"Professional" and "friendly" mean different things to different people. Describe tone through features a reader would notice:

• Formality — contractions or not, greeting style, technical vocabulary.
• Warmth — direct address ("you"), acknowledgement of the reader's situation.
• Confidence — hedged ("may", "it is possible") or direct.
• Energy — sentence length, exclamation marks (often best banned explicitly).

Example: "Tone: calm and reassuring; short sentences; no exclamation marks; acknowledge that the delay is frustrating before explaining the next step."

Supplying a short sample of the voice you want ("Match the tone of this paragraph:") is often more effective than adjectives.

Length: choose the right unit

Models handle structural limits more reliably than exact counts:

• Reliable: "3 bullet points", "one paragraph", "5 rows", "a heading and two sentences per section".
• Approximate: "about 150 words". Expect variation and check if it matters.
• If you need a hard limit (a character-limited field), verify it yourself or with a counter tool.

Structure for long outputs

For anything longer than a few paragraphs, give the skeleton: headings in order, what each section contains and how long it is. You can also ask the model to propose an outline first, approve or edit it, then ask for the full text section by section — a simple form of the decomposition you will learn on day 10.

Avoiding unnecessary instructions

Prompts tend to grow: every time something goes wrong, another rule gets added. Long prompts with many rules can:

• conflict ("be thorough" vs "be brief");
• dilute attention, so important rules are followed less reliably;
• make the prompt hard to maintain and test.

Trim with three questions:

1. Does this change the output? Remove filler such as "Please do your best" or "You are a helpful AI."
2. Is it already implied? "Write in English" is usually unnecessary if the prompt is in English.
3. Does it conflict with another rule? Resolve conflicts by stating a priority: "If you cannot meet both, prefer accuracy over brevity."

A useful test: delete one instruction, run the prompt a few times, and compare. If quality does not drop, leave it out.

Example: before and after trimming

Before (78 words): "You are a very helpful, friendly and knowledgeable AI assistant. Please do your very best. I would like you to write an email. The email should be polite. The email should be professional but also friendly. Please make sure the email is not too long and not too short. The email is to my landlord about the broken heating. Make sure to be polite. Thank you so much!"

After (38 words): "Write a polite email to my landlord reporting that the heating has not worked since Monday. Ask for a repair date this week. Tone: courteous and firm. Under 120 words, with a clear subject line."

The shorter prompt contains more useful information: the date, the request and a checkable length.