How to brief an email agent: a campaign plan you can actually review

“Write a great welcome email” leaves almost every consequential decision open. A better brief is: help a newly opted-in reader find the starter guide they requested, using one clear link and no unrelated sales pitch. The message now has a job. The agent can judge whether an opening sentence, image, or secondary button helps that job.

State the desired outcome separately from the metric. A click is an observable event; understanding the guide is not directly observable from that click. Choose a practical measure, explain its limitations, and keep complaints or unsubscribes visible alongside it. Do not tell the agent to maximize opens at any cost.

For this example, the deliverable is one draft, not a new lifecycle strategy. Name what is out of scope: changing signup consent, creating an offer, adding recipients, or enrolling anyone in a follow-up sequence.

Provide an approved source for the resource name, destination, sender identity, and any commercial terms. Attach a version or checked date where useful. Mark an assumption as an assumption. A previous campaign can illustrate tone, but its expired price should not become today’s source of truth.

Then define the permitted work. Reading an audience count, creating a draft, changing a template, and sending a campaign are different actions. Permission to do one does not imply permission to do the others. This matters whether your assistant runs in a chat client, Claude Code, Codex, or a custom tool runner.

Anthropic’s guide to effective agents describes using programmatic checks between workflow steps. Our application to email is deliberately concrete: drafting can be flexible; audience eligibility and send authorization need a separate, enforceable boundary.

This is a planning example. The names, limits, and review roles must come from your team’s policy before it becomes an execution instruction. A resource URL is required input, not something the model should guess.

Goal: deliver the requested starter guide to new subscribers.
Deliverable: one reviewable welcome-email draft; do not send.
Audience: confirmed marketing opt-ins awaiting this welcome email.
Exclude: unsubscribed, suppressed, already welcomed, or ineligible contacts.
Facts: use the supplied guide title, approved URL, and sender profile only.
Template: use the selected welcome template and retain its required footer.
Editable: subject, preview text, opening paragraph, and CTA label.
Protected: destination URL, consent rules, sender, and required variables.
Personalization: use a tested fallback when the first name is absent.
Timing: propose a schedule in the workspace’s explicit time zone.
Budget: propose one version and one revision; stop if required facts are missing.
Review: show audience counts, exclusions, links, assumptions, and changes.
Authority: draft creation only; sending requires a separate authorized action.
Success: verify the resource link and explain what the measured events mean.

A brief that says “exclude suppressed contacts” is not itself an enforcement mechanism. Your email system must still resolve the audience and check current eligibility at execution. A contact can unsubscribe after the preview was generated.

The review should contain enough context to make a decision without re-running the whole conversation. Ask for the chosen template and revision, subject and preview text, rendered email, plain-text version, destination links, and a list of unresolved variables. Put factual sources next to the claims they support.

For the audience, request counts with clear definitions: candidates considered, excluded by rule, eligible at preview time, and intended recipients. Overlapping exclusions should not be added as if they were unique people. If someone is both suppressed and already welcomed, the report needs a consistent counting method.

Request a difference summary after every revision. A change from “download the guide” to “claim your discount” is a commercial change, not just better wording. If the recipient set, protected URL, offer, sender, or schedule changes, the reviewer needs to see that before approving again.

A good result can be a blocked proposal. No approved resource URL, unclear consent state, a broken variable, or conflicting commercial terms should leave the draft waiting for clarification. Tell the agent who can answer, which evidence is missing, and which work can safely continue.

Give feedback in terms of the brief: shorten the opening, retain the resource destination, remove the second CTA. Do not approve a general intention and then let the agent silently transform the artifact. Review authority should refer to a specific version and intended action.

Before a later send, re-check eligibility and the applicable policy. A reviewable plan reduces ambiguity; it does not turn yesterday’s audience snapshot into today’s permission. Read The prompt is not the campaign, or begin with a Moosewave template and write down its protected parts.

• Anthropic: Building effective agents. Workflow patterns and programmatic checks. The campaign brief and review checklist here are Moosewave’s applied editorial example.

Sources checked 2 October 2026. Product behavior and documentation can change, so the linked primary source takes precedence if it differs from this article.