Content generation: copy, descriptions, emails

By default a model writes bland, watery prose: "in today's world", "it is worth noting", "let's take a closer look". You can spot such text instantly. To get a usable result, control three things, tone, structure, and length, and ban the clichés explicitly.

Tone is set by example, not by adjective

"Write in a friendly tone" barely works: every model has its own idea of friendliness. An example works:

system = """You write copy for an online store.
Tone, as in this example:
"This jacket shrugs off a downpour and doesn't restrict movement. The membrane breathes,
the hood adjusts, the pockets are right where your hands are."
Short, to the point, no bureaucratic language and no gushing adjectives."""

One or two samples in the system prompt set the style better than any description.

Set the structure firmly

If you need a specific form, define it as a list, not as a hope:

Email structure:
1. One sentence: the point.
2. Three bullets: what the customer gets.
3. One call to action.
No preamble and no half-page sign-off.

Mass-producing descriptions

For product cards, where you need hundreds of descriptions, run a cheap model in batches and pass the specifications in a structured form:

resp = client.chat.completions.create(
    model="qwen-flash",
    messages=[
        {"role": "system", "content": system},
        {"role": "user", "content": json.dumps(product, ensure_ascii=False)},
    ],
    temperature=0.7,
)

A temperature of 0.6 to 0.8 is a reasonable range: the texts vary but do not drift apart. For emails and important copy, lower it to 0.4 to keep the precision.

Choosing a model

• Mass descriptions, social media, drafts: Flash models. Cheap, fast, good enough.
• Landing pages, important emails, editorial guidelines: a stronger model (deepseek-pro, qwen3.8-27b). It holds a long structure better and hits the tone more precisely.
• Russian: test it separately. Some models slip into word-for-word calques from English. DeepSeek and Qwen write clean Russian.

Banning clichés

Add a stop list to the system prompt. It raises quality noticeably:

Forbidden: "in today's world", "it is worth noting", "it's no secret that",
"let's take a closer look", "in conclusion", "plays an important role".

Pitfalls

• Gushing. The model loves superlatives. Ask for "no evaluative epithets, only facts and benefits".
• Identical openings. In a batch of 200 descriptions, half will start with "This product...". Put a requirement in the prompt to vary the first sentence.
• Length drifts. The model interprets "write briefly" loosely. Specify it in characters or sentences: "no longer than 300 characters".

Do not drop the final human edit on important texts, but with a good system prompt the edits become "fix two words" instead of "rewrite everything".