Writing and content drafting

The task

You want a draft that fits the situation: an email, article, product update, documentation page, or social post. The goal is not “good writing” in the abstract. The goal is writing that sounds appropriate for the audience and does the job.

Why the old approach is outdated

Older writing prompts leaned on persona boilerplate: “You are a world-class copywriter with 20 years of experience.” Then they stacked tone adjectives on top.

That usually adds length without adding direction. Current models respond better to audience, format, constraints, examples, and revision criteria. A real style guide or two representative samples beat a paragraph of adjectives.

The current approach

Lead with the brief. Add a role only when domain context changes the answer.

Write a launch announcement email for our new API rate-limit dashboard.

Audience: existing API customers, technical but not necessarily engineers.
Length: under 200 words.
Goal: get readers to check their own usage.
Avoid: marketing superlatives, exclamation points, and a hard sales pitch.

Existing changelog entry:
{{changelog_text}}

The model has enough to make real choices: audience, length, outcome, boundaries, and source material.

Getting the voice right

Use samples when voice matters. Ask the model to match observable traits rather than invented personality.

This is the most practical use of few-shot prompting: examples teach the model the boundary that plain tone adjectives cannot.

Match the tone and sentence rhythm of these two examples of my writing.

<example>{{sample_1}}</example>
<example>{{sample_2}}</example>

Now draft a similar piece about:
{{new_topic}}

For editing, ask for a revision target instead of a rewrite command.

Revise this draft to sound more direct and less promotional. Keep the meaning and
specific facts. Do not add claims that are not already in the draft.

{{draft}}

What to avoid

• Long persona preambles in place of concrete constraints.
• Asking for multiple variants by default. Ask for one strong draft, then iterate.
• Saying only what not to do. Pair each constraint with the behavior you want instead.
• Asking the model to “make it sound human.” Give it samples, audience, and revision criteria.