05 · Controlling Tone, Length & Style

"Make it sound less robotic" is one of the most common requests people make of models — and one of the least effective. Style is controllable, but only when you describe it in terms the model can act on. This lesson turns taste into specifications.

Describe style along dimensions

Instead of one adjective, specify several concrete dimensions:

Dimension	Example setting
Formality	"Conversational; contractions are fine; no slang"
Sentence length	"Mostly short sentences; vary length; max ~25 words"
Person & address	"Second person ('you'); we = the company"
Vocabulary	"Everyday words; no buzzwords like 'leverage', 'synergy', 'seamless'"
Warmth	"Friendly but not effusive; no exclamation marks"
Structure	"Lead with the point; one idea per paragraph"
Confidence	"State facts plainly; hedge only where genuinely uncertain"

A compact style spec combining these is reusable across prompts and people.

Show a voice sample

Descriptions only go so far. A short sample of the target voice is often the strongest signal:

Write the product update in the same voice as the sample in <voice_sample>: match its
sentence length, level of formality and humour. Do not reuse its content or phrases.

<voice_sample>
We moved the export button. It was hiding in Settings, which, in hindsight, was a
strange place to keep it. It now lives at the top of every report.
</voice_sample>

The instruction not to reuse phrases matters; otherwise models tend to borrow distinctive lines from the sample.

Reading level and audience

Reading-level targets are useful when paired with an audience:

Rewrite for adults with no medical background, at roughly the reading level of a
general-interest newspaper. Keep all numbers and dosing instructions exactly as written.

Note the second sentence: when simplifying, specify what must not change.

Length, revisited

From Level 1: prefer structural limits. For longer pieces, set a length per section:

Intro: 2 sentences. Three sections of ~80 words each, with a one-line heading.
No conclusion section.

When a model keeps writing too much, the cause is often content scope, not wording: tell it what to leave out ("skip background; the reader knows the product").

Avoiding the "assistant voice"

Many people recognize a generic AI register: stock openings, inflated adjectives, forced summaries, lists of three everywhere. To steer away from it:

• Ban specific phrases you dislike and describe the replacement behaviour.
• Ask for concrete details from the provided material rather than general claims.
• Ask for the draft to open with the most specific fact available.
• Provide a real sample of writing you like.

Banned-phrase lists work best when short and specific; a very long list can itself become a distraction.

Measuring style with code

Some style properties are measurable. This script reports sentence lengths, exclamation marks, and banned words — useful when you're comparing prompt versions across many outputs:

import re

BANNED = {"leverage", "seamless", "delve", "synergy", "game-changer"}

def style_report(text: str) -> dict:
    sentences = [s for s in re.split(r"(?<=[.!?])\s+", text.strip()) if s]
    words = re.findall(r"[A-Za-z'-]+", text.lower())
    lengths = [len(re.findall(r"[A-Za-z'-]+", s)) for s in sentences]
    return {
        "sentences": len(sentences),
        "avg_words_per_sentence": round(sum(lengths) / len(lengths), 1),
        "longest_sentence": max(lengths),
        "exclamations": text.count("!"),
        "banned_found": sorted(BANNED & set(words)),
    }

draft = ("Our new dashboard lets you leverage real-time insights! "
         "It is fast. Reports load in under two seconds, even for large accounts, "
         "and you can share them with one click.")
print(style_report(draft))

Output:

{'sentences': 3, 'avg_words_per_sentence': 9.7, 'longest_sentence': 18, 'exclamations': 1, 'banned_found': ['leverage']}

Metrics like these don't capture quality, but they turn vague complaints ("too salesy") into something you can track across versions.

Worked example: one message, three audiences

Facts: scheduled maintenance, Saturday 02:00–04:00 UTC, the app will be read-only, no data will be lost.

Write three versions of a maintenance notice using only the facts in <facts>:
1. In-app banner: max 20 words, neutral.
2. Email to business customers: 60–80 words, formal, includes the times in UTC and
   a suggestion to plan exports beforehand.
3. Status-page entry: one factual line with times; no reassurance language.
Label each version with its number.

Specifying each audience's constraints in one place lets you compare the variants side by side and keeps facts consistent across them.

How It Actually Works

Style is distributed across every token choice — word frequency, sentence boundaries, punctuation — so it is shaped by everything in the context: instructions, samples, even the style of your own prompt. A prompt written in terse, plain sentences tends to pull the output that way; a rambling prompt pulls the other way. Voice samples work particularly well because they supply the dense, low-level statistical cues (rhythm, word choice) that an adjective like "punchy" only gestures at.

The default "assistant voice" is itself a learned style: fine-tuning on preferred responses creates recognizable habits, and without strong cues the model falls back on them. Explicit specs and samples give it stronger cues to follow instead.

Common mistakes

• Single-adjective style instructions ("make it engaging").
• Voice samples without a "don't copy content" instruction.
• Simplifying without protecting facts like numbers and names.
• Huge banned-word lists instead of describing the target.
• Writing the prompt in the opposite style of the output you want.

Exercise

1. Write a style spec (6–8 lines, using the dimensions table) for a real voice you need: your team's docs, your newsletter, your own emails.
2. Find a 60–100 word sample that exemplifies it.
3. Generate the same piece three ways: no style guidance; spec only; spec + sample.
4. Run style_report (or your own checks) on each and read them aloud. Which is closest, and which dimension is still off?