Prompt steps: let Claude write each email

How to do it

1. First add your Claude API key in Settings → Integrations (see "Integrations: Claude API key"). Claude writes with your own Anthropic account, so Prompt can't be picked until a key is saved.
2. In the campaign builder, select an Automated email or Manual email step and, under "Write this email from", switch Template to Prompt. If the step already has template text or A/B variants, you're told first that they're removed when it saves: a prompt step has neither.
3. Write the prompt: who you are, what you offer, the angle, the tone, how long the email should be and the call to action. Click a variable chip to add {{first_name}}, {{company}} or any column from your list; Claude fills each one from that person's details. Up to 4,000 characters. The rules every email keeps to are listed under the box, and "Variables this prompt uses" shows what happens to someone with one of them empty.
4. Pick the Model. Claude Sonnet 5 is the default (balanced quality and cost); Claude Haiku 4.5 is the fastest and cheapest; Claude Opus 5.5 (highest quality, slower) and Claude Fable 5.1 (most capable) cost more. The rough cost per 1,000 emails for the writing shows under the picker; the second read that checks each email adds about $3 per 1,000.
5. Choose the Subject: "Claude writes it" ("Claude writes a new subject" on a follow-up; step 1's default), "Fixed subject" (the same subject for everyone, variables allowed) or, on a follow-up, "Same thread (Re: …)": no new subject, it follows up in the same thread as the earlier email (the default for later steps).
6. Click Write a Preview to see the email for the person picked in the preview (or a sample person if nobody is enrolled yet), with its tokens and roughly what it cost, and any suggestions. Send test writes one for the first enrolled person with every variable filled in (or a sample person) and sends it to the address you type.
7. Click Check Prompt. Claude writes the step for up to 3 people in the campaign who have every variable filled in, through every check a real email goes through, and shows you each email with ✓ or ✗. The step then reads Checked ✓ (or Failed); a new or edited prompt reads Needs a check. Check Prompt checks the saved prompt, and the builder saves a few seconds after you stop typing.
8. Launch the campaign as usual once every prompt step reads Checked ✓. Each person's email is written shortly before it's due and goes out within the campaign's schedule like any other email.

Good to know

• Costs are billed by Anthropic to your account, not by MailFleet. Rough estimates for a typical short email (longer prompts, longer emails and follow-ups cost more): about $3.50 per 1,000 emails on Claude Haiku 4.5, $7 on Claude Sonnet 5, $22 on Claude Opus 5.5 and $55 on Claude Fable 5.1, plus about $3 per 1,000 for the second read that checks each email. Previews, test emails, Check Prompt (up to 3 emails each time), rewrites and retries are billed the same way. MailFleet adds no charge and never uses its own Anthropic account for your emails.
• The rules every email keeps to, whatever the prompt asks: only the email to the person, with no notes to you, no offers to shorten or change it, no alternatives, no placeholders, and never written as an AI or saying it was written by AI (writing about AI products your prompt is about is fine); at least about 15 words; plain paragraphs, no bullet points or headings; no made-up numbers, clients or case studies about you (saying in general what you do and the value you bring is fine, and you don't need success stories); and no links, web addresses or email addresses unless your prompt gives them word for word.
• Before an email can be used it passes automatic checks and a second read by Claude Haiku 4.5 on your key. Only a broken email is stopped: one that breaks a rule above (for example a link that isn't in your prompt, bullet points, or fewer than about 15 words), text meant for you rather than the person, a placeholder or an unfilled bit, the wrong name or company, made-up specifics (an event or number about the person that isn't in their details, or stats, named clients, case studies, prices or guarantees about you that aren't in your prompt), the email talking about itself as AI, or an email that's cut off. A stopped email is never sent; it's written again.
• Anything else the second read notices (generic or vague wording, no case studies or results, general claims about what you do, a thin point or a weak call to action) shows as Suggestions under the preview and under each Check Prompt email. Suggestions never stop an email.
• Every prompt step must pass Check Prompt before the campaign can launch, whether you launch it in the app, over the API or from Claude. Changing the prompt, its model, its subject setting or its signature setting needs a new check. Check Prompt also fails when a {{variable}} isn't a field of the campaign's leads (check the spelling), when the prompt uses an email address or phone number variable (Claude is never given those), when the campaign has no leads yet, or when a same-thread step has no email before it to follow up on.
• Empty variables: a person with a {{variable}} your prompt (or your fixed subject) uses empty, and no fallback like {{first_name|there}}, gets nothing written. On an automated step they skip the step, just as a template skips a missing variable; on a manual step their task opens empty for you to write. A same-thread follow-up for someone the campaign hasn't emailed yet gets nothing written either: on an automated step they skip it, and on a manual step their task opens empty for you to write.
• Add the mailbox signature: on, your mailbox signature goes below Claude's email; off, Claude ends with a short sign-off, without a name or details unless your prompt gives them. On an automated step, emails Claude writes send exactly like your other campaign emails: same mailboxes, schedule, daily limits and email credits, with the opt-out line and threading added by MailFleet. On a manual step you send it from Tasks like any manual email task.
• If Claude can't write someone's email yet (Anthropic is busy, or the answer was stopped), that person waits on the step and it's tried again after 1 minute, then 5, then 15, then every hour. Nothing unwritten is ever sent. After three failed tries for someone you're notified (Notifications → Campaigns, at most once a day per step).
• If Claude's answers for someone keep failing the checks (5 times for the same prompt), or Claude says it can't write their email from the prompt and their details, it stops trying for them so you don't pay for the same answer again. They wait and nothing is sent. The step shows "Claude stopped on N people at this step" with a button, "Retry the N people Claude stopped on"; or fix the prompt, and once the edited prompt passes Check Prompt they're written again.
• If Anthropic refuses your key (deleted, out of credit, or not allowed to write), writing stops for the whole workspace and you're told at once. Save a working key in Settings → Integrations, or add credit at console.anthropic.com: MailFleet checks a refused key again every hour and starts writing on its own once Anthropic accepts it.
• If your Anthropic account can't use the model you picked, you're told at once. Pick another model on the step and run Check Prompt again; the people waiting are then written with it.
• Editing a prompt step on a running campaign: the people on that step wait until the edited prompt passes Check Prompt, then their emails are written from it (any written earlier but not yet sent are written again). Emails already sent don't change.
• Manual email steps: Claude writes each task's email up to a day before it's due, or at once when you click Write It Now in Tasks. The task shows "Claude is writing this email…", then the email with Edit email to change anything and Rewrite with Claude for a new version from any model. Write it yourself is there whatever Claude is doing, and nothing sends until you click Send & next. See "Send a manual email task".
• Prompt steps work on automated and manual email steps in Email campaigns. They aren't available in subsequences, Multi Channel or LinkedIn campaigns, or on call steps, and a prompt step has no A/B variants.
• What Claude is given for each email: your prompt; the person's name, job title, company, industry, location, company size, website, LinkedIn profile address and time zone; all of your list's extra columns (not ones holding email addresses or phone numbers); and, for a follow-up, the earlier emails of this sequence to them. Never their email address or phone number, and nothing read from your mailbox. It goes through your own Anthropic account. See the Privacy Policy, section 04.
• The REST API and the Claude connector take prompt steps too: on an email or manual step set mode "prompt" with a prompt, and optionally model and subject_mode (subjectMode on the connector). Starting such a campaign needs a working key and every prompt step passing Check Prompt, which you run in the campaign builder.