Good drafting prompts follow the same discipline as good delegation. When you hand a writing task to a colleague, you tell them who it's for, what it must achieve, what facts to use, and what tone fits. Anthropic's own prompting guidance formalizes this as its golden rule: treat Claude like a brilliant but brand-new employee — if a person with minimal context would be confused by your brief, Claude will be too. A vague brief in, a vague draft out.
The brief-in, draft-out structure
A reliable drafting prompt has four labeled parts. Anthropic's best practices recommend wrapping each in its own descriptive XML tag — consistent tags help Claude keep instructions, facts, and inputs apart:
| Section | What goes in it | Example |
|---|---|---|
<audience> | Who reads it, what they know, what they decide | "CFO; skeptical of the project; decides budget renewal" |
<goal> | The single outcome the document should produce | "Secure sign-off on the Q3 extension" |
<facts> | The raw material: figures, dates, decisions, constraints | Bullet list from your notes — the only facts Claude may use |
<format> | Length, structure, tone, channel | "One-page memo, three headed sections, plain and direct" |
The <facts> section is the anti-hallucination device. Pair it with an explicit rule: "Use only the facts provided; if something needed is missing, insert a [TODO] marker rather than inventing it." That single line converts silent fabrication into a visible gap a human can fill.
Set the voice once, in the system prompt
For a drafting tool used across a team, put the durable identity in the system prompt and keep per-request briefs short. Per the official guidance, giving Claude a role via the system prompt — even a single sentence — focuses behavior and tone: "You are a communications writer at a logistics company; you write plainly, lead with the decision needed, and never use marketing superlatives."
Two refinements from Anthropic's best practices make role prompts markedly better. First, explain why a rule exists so Claude can generalize it — "avoid idioms, because many readers are not native English speakers" travels further than "avoid idioms." Second, ask explicitly for what you want above the baseline: the docs advise requesting "above and beyond" behavior directly rather than hoping the model infers it. If you want the draft to anticipate objections, say "anticipate the two strongest objections and address them."
Examples beat adjectives for tone
"Professional but warm" means something different in every company. The higher-bandwidth pattern is few-shot: include 3–5 relevant, diverse examples of past documents in the house voice, wrapped in <example> tags, per the official guidance. Diversity matters — one short email, one longer memo, one delicate message — so Claude learns the voice rather than the template. A brief note on each ("note how this one delivers bad news without hedging") sharpens the signal further.
Draft, then revise — in separate turns
Resist the mega-prompt that asks for a perfect final document in one shot. Drafting workflows do better in two passes: generate a draft from the brief, then send revision instructions in a follow-up turn ("tighten section two; the CFO cares about cash impact, not architecture"). This matches how humans edit, keeps each instruction set small and unambiguous, and gives reviewers a natural checkpoint. For flows where a second pass should critique rather than rewrite, see the companion piece on review and critique prompts.
One production note: length instructions belong in the prompt ("at most 400 words"), while max_tokens is only the safety ceiling. And before rolling a drafting prompt out to a department, run it against a handful of real briefs from that team — Anthropic's prompt-engineering overview puts clear success criteria and empirical tests ahead of prompt polish for good reason.
Where to go next
Deepen the voice work with tone and style prompting and role prompting. To turn a good drafting prompt into a governed team asset, see prompt template design and prompt versioning.