Prompt Engineering

How to Write Claude Email Prompts That Don't Sound Like a Robot

Published April 19, 2026 · 7 min read

You open a new Claude chat. You type "Write a welcome email for my SaaS." You hit enter. Thirty seconds later Claude hands you back something that starts with "We're thrilled to have you on board!" and ends with a CTA button that says "Get Started Today!"

It's fine. It's also identical to every welcome email you've deleted unread for the last six years.

The problem isn't Claude. Claude is doing exactly what you asked — producing an average email for an average SaaS, because your prompt was 9 words long and told it nothing about you, your reader, or what "welcome" is supposed to do in your specific business.

Generic prompts get generic outputs. That isn't a model limitation. It's an input problem. And the fix is the same whether you're writing welcome emails, cold outreach, renewal nudges, or internal updates: give Claude five specific pieces of information, every time, before you ask for copy.

Why one-line prompts produce corporate filler

Large language models are pattern matchers across everything they've ever read. When you give Claude almost no context, it falls back on the statistical average of "welcome email" in its training data. That average is drowning in SaaS onboarding templates from 2018–2024 — all written by marketing teams using the same three or four playbooks.

So you get the average of the average: safe, forgettable, vaguely corporate. "Hope this email finds you well." "Excited to have you." "Questions? Just reply!"

Claude can write wildly better than that. You just have to close the gap between what you know about the situation and what the prompt actually tells the model.

The 5-part structure

Every high-signal email prompt I write has five parts. Skip one and quality drops a full point. Include all five and the first pass is usually shippable with minor edits.

1. Reader state — who this email is going to, and where they are mentally right now
2. Goal — the single outcome you want from the email
3. Voice — how it should sound (three words or two example texts)
4. Constraint — length, format, forbidden phrases, non-negotiables
5. Example — one email (yours or a competitor's) that already sounds right

Let's go through each one with the specificity that actually makes the difference.

1. Reader state

Claude can't guess who your reader is. "Customer" means nothing. "Lead" means nothing. You have to describe the person's current situation in one or two concrete sentences.

Just signed up 3 days ago. Hasn't logged in yet. Paid the $49 Pro plan, so they're serious enough to spend money but haven't seen value yet.

That single paragraph changes the email more than any adjective you could throw at it. Claude now knows the reader paid (no discount pitch needed), hasn't logged in (onboarding friction is the real problem), and is on the Pro plan (the email shouldn't sound like it's for a free-tier user).

2. Goal

One outcome per email. Not three. If you write "welcome them, get them to set up, and introduce the blog," Claude will try to cram three CTAs into one message, and the email will feel bloated because it is bloated.

Pick the single next action. "Get them to log in and complete setup in the next 48 hours." That's it. Everything else is a follow-up email.

3. Voice

Voice is where most prompts fail, because adjectives are ambiguous. "Casual" means different things in SaaS Twitter, in luxury brand copy, and in a fintech newsletter. Claude's training data includes all three. It picks the average. You get LinkedIn voice.

You have two better moves:

• Three-word voice description. Instead of "casual," say "warm, direct, dry." Instead of "professional," say "confident, specific, unhedged."
• Two-example voice match. Paste two emails (or two tweets, or two paragraphs from your site) that already sound right, and say "match this energy." Examples beat adjectives by a factor of ten.

If you have examples, always use them. Adjectives are a fallback.

4. Constraint

Without constraints, Claude writes to the length of its training data — and email training data skews long, padded, and overly polite. Constraints are the part most people skip, and they're the part that most improves the output.

Useful constraints:

• Word count ceiling — "under 120 words"
• Paragraph count — "three paragraphs max"
• Forbidden phrases — "Hope this finds you well", "circle back", "just wanted to"
• Structural rules — "no greeting, no sign-off, start with the hook"
• CTA rules — "one CTA, as a P.S., not a button"

Forbidden-phrase lists are disproportionately powerful. If you've ever wondered why Claude defaults to "I hope this finds you well," the answer is it's in its training data tens of millions of times. Explicitly banning it forces the model to write something else — and the something else is almost always more interesting.

5. Example

This is the single highest-leverage move in the whole structure. Paste one email — yours, a competitor's, or just something you think sounds right — and tell Claude "match this structure and energy."

Why it works: Claude is infinitely better at imitation than at following abstract style rules. An example gives it structure, rhythm, sentence length, paragraph breaks, tone, vocabulary, and CTA placement — all at once, concretely, with zero ambiguity.

If you only have time for one upgrade to your prompts, add the example. It alone closes most of the gap between "generic output" and "this is shippable."

Before and after

The vague prompt

Write a welcome email for my SaaS.

Output (paraphrased): "Welcome to [Product]! We're thrilled to have you on board. Our platform is designed to help you achieve your goals faster than ever before..." — you know the rest.

The 5-part prompt

Write a welcome email.

Reader: Just signed up 3 days ago for the $49 Pro plan. Hasn't
logged in yet. Solo founder, building a client-reporting tool
for agencies.

Goal: Get them to log in and connect their first client
workspace in the next 48 hours.

Voice: Warm, direct, low-bullshit. No "hope this finds you well."
Match the tone of these two emails:
[paste email 1]
[paste email 2]

Constraint: Under 120 words. No greeting, no sign-off. One CTA.
Forbidden phrases: "excited", "onboard", "journey", "seamless".

Example structure to match:
[paste one email that already works for you]

The output from that prompt is not a polished corporate welcome. It's a three-paragraph message that names the specific friction (you haven't logged in yet), makes one ask (connect your first workspace), and sounds like it was written by a human who has actually used a client-reporting tool.

Same model. Same topic. The difference is you gave Claude something to work with.

Where this structure applies

Everywhere you write email with an LLM. The five parts don't change:

• Cold outreach — Reader state: what do you know about this prospect's business? Goal: book a call, or get a reply? Voice: your actual sales voice, with two examples. Constraint: no templated openers.
• Re-engagement — Reader state: went dark when, why do you think? Goal: one small action, not "re-engage."
• Internal updates — Reader state: what does the team already know? Goal: one decision or one piece of awareness.
• Renewal nudges — Reader state: are they a heavy user, a light user, or a ghost? The email should be radically different for each.

The structure scales up, too. When you run emails through an agent — a system prompt that handles voice, constraints, and example-matching automatically on every invocation — the 5-part framework becomes the skeleton of the agent itself.

A checklist you can save

Before you hit enter on any Claude email prompt, ask:

1. Did I describe the reader's current state in one specific sentence?
2. Did I name one outcome — and only one?
3. Did I define voice with either three concrete words or two example texts?
4. Did I list at least two constraints (length, forbidden phrases, structural rules)?
5. Did I paste one example email that already sounds right?

If any answer is "no," add it before you send. Two minutes of prompt upgrade saves twenty minutes of editing bad output.

The deeper point

Email isn't the only format where this matters. Every prompt in every tool I've built — for landing pages, proposals, sales scripts, reel hooks, long-form blog posts — has some version of these five parts. The language changes. The skeleton doesn't.

Claude isn't a magic content generator. It's a precision instrument that performs at the level of the specificity you hand it. Give it ambiguity, get averages. Give it concrete inputs, get concrete outputs.

Every email you send is a chance to stop sounding like the average welcome email from 2019. It takes five fields.