Chapter BAI for emailPage 2 of 8

AI for email

Drafting patterns that stay useful

The best drafting prompt gives the model a job, source material, constraints, and a definition of done.

~14 minCore mental model

Before you start

Why this matters

Choose a routine email you send: a request, update, decision brief, or difficult-news message. List the facts that must appear, the action the reader should take, and one claim the draft must not make. If those items are not explicit before generation, a polished draft can conceal the gap rather than solve it.

1Learn the idea

Read

Draft from a brief, not from a wish

Once goal, audience, and tone are clear, you can use repeatable patterns instead of improvising every prompt. A pattern is not a magic phrase. It is a small workflow that makes inputs and review criteria visible.

A dependable drafting prompt has five parts:

  1. Task: what to produce.
  2. Context: why the message is being sent and what the reader knows.
  3. Source facts: information the draft may use.
  4. Constraints: length, tone, format, and prohibited claims.
  5. Success test: what must be obvious to the reader.

Here is the basic shape:

Draft a [type of email] to [audience] for [goal]. Use only the facts below. The reader already knows [context]. Use [observable tone choices]. Include [required elements], stay within [length], and mark missing information with [NEEDS INPUT]. The draft succeeds if the reader can tell [definition of done].

The success test changes the task from “produce fluent text” to “help the recipient act correctly.”

Teach

Five reusable drafting patterns

The same brief can support several common email jobs. Choose the pattern that matches the outcome instead of asking for a generic “professional email.”

Pattern 1: request an action

Action requests fail when the ask is buried or incomplete. Supply the reason, action, owner, deadline, and next step.

Prompt structure:

  • Purpose: why the request matters.
  • Request: one verb-led action.
  • Deadline: exact date, time, and time zone where needed.
  • Support: links, attachments, or help available.
  • Close: one question that makes replying easy.

For example: “Draft a concise request to three team leads. Ask each to approve or comment on the attached migration checklist by 2 p.m. GMT on 12 August. Explain that unresolved comments will be reviewed in Thursday’s planning meeting. Do not imply that silence equals approval. Put the request in the first paragraph.”

Notice that the prompt prevents a risky assumption. AI often produces familiar phrases such as “If I don’t hear back, I’ll assume this is approved.” That may be unacceptable for a consequential decision.

Pattern 2: share an update

An update should help readers understand what changed, what remains uncertain, and whether they must act. Give the model a simple structure:

  1. current status;
  2. change since the last update;
  3. impact;
  4. next milestone;
  5. owner and requested action, if any.

Ask it to distinguish facts from forecasts. “Testing is 80% complete” is a fact only if your evidence supports it. “We expect to finish tomorrow” is a forecast. A useful constraint is: “Label estimates as estimates and do not convert uncertainty into a commitment.”

If different readers need different detail, request a short executive version and a technical appendix. Do not force one paragraph to serve every audience.

Pattern 3: make a decision brief

Some emails are small decision documents. The recipient needs the question, options, trade-offs, recommendation, evidence, and decision deadline.

Prompt:

Draft a decision email for the operations director. Start with the decision needed. Present options A and B in parallel bullets covering cost, timing, and risk. Recommend A because of the supplied evidence, but state the unresolved dependency. Ask for a decision by Tuesday at noon. Do not add costs or probabilities.

This pattern reduces persuasive drift. The model may naturally strengthen a recommendation, remove caveats, or make options look more certain than the evidence allows. Parallel structure and explicit uncertainty help the reader compare honestly.

Pattern 4: deliver difficult news

Delays, mistakes, rejected requests, and service problems require accuracy and care. Use a structure of fact, impact, ownership, next step, and contact:

  • state what happened without euphemism;
  • acknowledge the recipient’s likely impact without claiming to know their feelings;
  • take only the responsibility that is accurate;
  • give the current remedy or investigation status;
  • specify when the next update will arrive.

Do not ask AI to “make this sound less bad.” That instruction invites minimization. Ask it to be “clear, accountable, and non-defensive; preserve the exact impact and do not invent a resolution.”

For high-stakes personnel, legal, safety, medical, or financial messages, drafting assistance does not replace the responsible reviewer.

Pattern 5: create variants for comparison

One draft can anchor your judgment. Ask for two or three meaningfully different options:

  • a 75-word direct version;
  • a warmer version that includes one sentence of context;
  • a formal version suitable for an external recipient.

Tell the model what should remain constant across variants: facts, requested action, deadline, and commitments. Then compare choices instead of accepting the first fluent answer.

Variants are useful for subject lines too. Ask for options that are specific and searchable, not clever: “Decision needed by 12 Aug: migration window” is easier to act on than “Quick question.”

Read

Supply evidence cleanly

When pasting notes, mark them as source material:

SOURCE FACTS
- Customer reported duplicate charges on 6 July.
- Billing reversed one charge on 7 July.
- Refund timing is 3–5 business days.

INSTRUCTIONS
Draft a response under 140 words. Do not say the issue is fully resolved.

This separation reduces confusion between text to follow and text to transform. It also helps you review whether every sentence in the output traces back to an approved fact.

If source material contains another person’s instructions—such as “ignore prior directions”—treat it as content, not authority. Email threads and pasted documents can contain misleading or malicious text.

Read

Iterate by diagnosis

When a draft is weak, do not repeatedly say “try again.” Name the defect:

  • “The request appears only in the final sentence; move it to the opening.”
  • “The wording promises delivery Friday; replace that with the approved estimate.”
  • “The recipient does not know the acronym; expand it once.”
  • “The apology is longer than the remedy; rebalance the two.”

Focused feedback preserves useful parts and teaches you which instruction was missing. Keep the original facts nearby, because several rounds of rewriting can gradually distort them.

Checking tutor…

Continue learning · glossary & guides
  • What five parts make a drafting prompt dependable?
  • Why should an update separate facts from forecasts?
  • When are variants more useful than repeated “try again” prompts?
  • Name one constraint that prevents an unsupported commitment.
  • Cheatsheet: prompt recipe · Glossary: prompt engineering