Email Marketing Bible

Email Marketing Bible, Skill Reference

> Source: EMB (17 chapters, 4 appendices). Full guide: https://emailmarketingskill.com > Built from 908 sources and the experience of running SmartrMail (~28,000 customers, 6B emails, sold 2022). > Two parts. Part A is the operating manual: read it when you are acting (building a flow, sending, diagnosing, designing). Part B is the dense reference: drop into it for facts, frameworks, and benchmarks. > Benchmarks are as of mid-2026. Verify time-sensitive figures (inbox rules, ESP features, pricing) before acting on them. > Section index with full-chapter links is at the bottom.

---

PART A: OPERATING MANUAL

0. AGENT OPERATING RULES

You may be driving a real ESP through MCP or connectors. You can create segments, draft copy, compose campaigns, build flows, and stage sends. Treat every one of those as a live action with consequences.

Send safety (hard gates, never skip):

• Never send or schedule to more than one recipient without explicit human approval in this conversation ("send it" or equivalent). Single-recipient test sends still need a yes.
• Always dry-run / preview first. Surface the preview URL before asking for approval.
• Always show the approval packet before any send: audience size, exclusions/suppressions applied, subject, preview text, send time, sender identity (from-name + reply-to), unsubscribe present (yes/no), and any compliance risk.
• Block the send if: authentication is missing, the unsubscribe or physical address is absent, the complaint rate is at or above 0.1%, the consent basis is unclear, or the audience includes suppressed/bounced/complained contacts.
• Never probe unknown mutating endpoints on a live audience. Anything with /send, /dispatch, /trigger, /fire, /publish in the path can dispatch immediately. If you cannot find the documented "approve scheduled" path, ask the human to click it. Use sandbox or cloned campaigns with seed lists when testing behaviour.
• Separate the modes. Transactional, marketing, lifecycle, and cold outbound have different rules, domains, and consent bases. Never mix them.
• Log what you do. State every autonomous action you took (segment changed, flow edited, campaign created, send staged) so the human can audit it.

When to use this skill: building/auditing an email programme, designing flows or copy, diagnosing deliverability, choosing a platform, pulling a benchmark, or operating an ESP from an agent. When not to: it is not a substitute for the human's final send decision, brand voice, or legal sign-off.

1. TASK ROUTER

Map the request to a procedure. One hop.

| Intent | Go to | You need | You produce | |---|---|---|---| | Audit an email programme | §2 loop, then the relevant reference | account access (read), recent sends | gap list + priorities | | Build / prioritise a flow | §7 Flow Recipes + §2 | business model, trigger, audience, offer, exclusions | flow spec + copy brief | | Send a campaign | §3 Pre-Send Checklist | list/segment, consent basis, copy, sender, timing | approval packet | | Diagnose deliverability | §11 Deliverability Triage | domain, ESP, bounce + complaint rate, recent changes | severity + remediation | | Write or de-slop copy | §4 Anti-Slop Copy | audience, offer, brand voice, one real proof | revised copy + rationale | | Design an email | §5 AI Design + §16 Decision Table | brand kit/tokens, archetype, goal | template brief → MJML/React Email | | Pick a platform | §13 Platform Selection | list size, use case, stack, budget, agent-driven? | shortlist + tradeoffs | | Pull a benchmark | Appendix | industry, email type | figure + caveat | | Cold outbound | §14 Cold Email | offer, ICP, sending domains, volume | sequence + infra plan |

2. AI EMAIL AUTOMATION (the operating model)

Mid-2026: the marketer's job moved from operator to director. You do not click the campaign together; you brief an agent, govern it, and own the send button. Klaviyo (Composer), HubSpot, Mailchimp, Customer.io, Brevo, beehiiv and nitrosend all ship prompt-to-campaign agents now, all with a human-approval gate as the default.

The loop: read state → reason → act → verify. Do not start by composing. Start by reading the account: lists, flows, recent campaigns, deliverability, suppressions. The single most useful first move with an MCP-connected ESP is "audit my account and tell me what is missing." Then reason, act on one thing, verify the result.

Automate vs keep human:

• Safe to automate: send-time optimisation, subject-line variant generation + A/B testing, cart/browse-abandonment triggers, post-purchase cross-sell recs, first-draft copy.
• Keep human: editorial/brand voice, strategy (segment priority, flow sequencing), creative direction, deliverability/domain management, and the final send.

The autonomy dial. Run "ask mode" (confirm before each action) by default; only move toward autonomous execution on narrow, reversible, low-brand-risk tasks, and keep an undo. Roll out read-only access (analytics) before write access (drafting, segments, sends).

Controlling an ESP from AI, both surfaces. This is not Anthropic-only. ESPs ship MCP servers (Klaviyo, Resend, Mailgun, beehiiv, MailerLite, Omnisend) and apps inside both Claude and ChatGPT (Mailchimp, Omnisend). MCP is becoming a cross-vendor standard. When advising, cover the surface the user actually uses.

Pre-send preflight (run before any stage): resolve tracking-wrapped CTA URLs to their real destination (links are wrapped at render, so "does the button point to the right URL" needs the decoded target, not the wrapper); spam score; image-to-text weight; dark-mode lint; and a test send to a seed address. Resend, nitrosend and others build versions of this in; if the tool does not, do it yourself.

Silent failure is the real risk. The worst outcome is a flow that quietly stops sending or whose analytics go missing, caught days later. Recommend a recurring AI health digest: which flows have not fired, which are erroring, which metrics dropped. You cannot catch silent failure without a scheduled review.

3. PRE-SEND CHECKLIST

Before staging any send, confirm every line. Surface the result to the human, then wait for "send it".

• [ ] Audience: size, segment logic verified against actual counts (AI-generated segments can be over-broad)
• [ ] Suppressions applied: unsubscribed, bounced, complained, globally suppressed, recently-emailed (frequency cap), active support issue
• [ ] Authentication: SPF, DKIM, DMARC aligned and at p=quarantine or stronger (required by Outlook for 5K+/day)
• [ ] Unsubscribe present (one-click, RFC 8058) + physical address present
• [ ] Copy: anti-slop pass done (§4), one clear CTA, subject ≤45 chars, preview adds info
• [ ] Design: mobile single-column ≤600px, dark-mode safe, alt text, live-text headline (§5)
• [ ] Links resolve (decode wrapped CTAs), no broken/placeholder URLs
• [ ] Sender identity correct (from-name + monitored reply-to), right account/brand
• [ ] Send time set; consent basis valid for this audience and content type
• [ ] Test send reviewed in a real inbox with real merge data
• [ ] Human approval captured

4. ANTI-SLOP COPY PROTOCOL

Raw LLM copy is now a deliverability liability, not just a quality one. Google filters high-AI-similarity text harder, consumers trust a brand less when they spot AI, and at scale the tell is structural sameness. Agents draft; humans and brand voice finish.

• The deepest tell is the absence of stakes. AI is perfectly balanced and takes no position. Inject one genuine, defensible opinion per email. Ask the draft: where do I actually disagree with this, where is it too safe.
• Burstiness. AI holds one rhythm. Humans vary it. Alternate long and short sentences; put a hard short line (3-5 words) after a long one, at least once per section.
• Blacklist (lint before send): delve, leverage, foster, ignite, empower, unleash, streamline, navigate, seamless, robust, cutting-edge, transformative, multifaceted, pivotal, dynamic, comprehensive, tapestry, landscape, beacon, realm, journey, "furthermore", "moreover", "in today's fast-paced...", "I hope this email finds you well".
• Syntax fingerprints (survive find-and-replace): "it's not X, it's Y" (cap once per email), rule-of-three adjective padding, copula avoidance ("serves as / stands as" instead of "is").
• Specificity is the cheapest humaniser. Real numbers, names, dates beat abstractions. Pull one real metric or fact from the brand's own data into every email.
• The em dash is overused by LLMs but not a reliable tell alone. The EMB no-em-dash-in-body rule is a brand-voice choice, not detector evasion.
• Workflow: human strategy → AI draft → human edit. For high-personality formats (founder letter, welcome), write rough human thoughts first, then have AI tighten. Never AI-first.

5. AI EMAIL DESIGN PROTOCOL

The 2026 design risk is not ugly emails, it is forgettable ones. AI defaults to competent and generic. Force it off its defaults.

• Context beats prompt. The biggest lever is not better wording, it is the design context the agent can read: a brand kit, design tokens, a components/exemplars set, a rules file. Feed those before iterating on prompts. Generic input produces generic output.
• Generate into a safe substrate, not raw HTML. Have the agent emit MJML, React Email, or Maizzle, which compile to inbox-safe HTML and handle Outlook. Raw-HTML-from-a-prompt is the classic slop trap.
• Anti-slop design rules: own one colour (run it 30-60% of the surface; colour drives ~80% of brand recognition); restraint beats decoration (AI over-decorates, the job is subtraction); real photography/captures, never AI-stock (visible AI imagery lowers trust); bold live-text headlines, never image-based (accessibility, dark mode, and so Gemini can summarise); single message, strong negative space. Ban the AI-default purple-to-blue gradient and beige washes.
• Compliant by default. Bake the constraints into the prompt: single column ≤600px, 44px tap targets, role="presentation" tables, dark-mode-safe colours (never pure #000 bg or #fff logos, use ~#121212), alt text on every image.
• QA before stage: mobile render, dark-mode, image weight (0.5% |

| Bounce rate | 3% | | Spam complaint rate | 0.3% | | List growth rate | 3-5%/mo | 5%+/mo | Negative | | Inbox placement | 85-94% | 94%+ | 12% CTOR) · cold → positive reply (3-5%) · newsletter → clicks/replies (opens are noise).

• Attribution: U-shaped (40/40/20) to start; incrementality is the gold standard. Well-run ecommerce: email drives 25-40% of revenue.
• Ask your data with AI. Query live ESP data conversationally via MCP/connectors instead of building dashboards; use AI for anomaly flagging and A/B readouts. Pair with the health digest (§2).
• Send frequency: track revenue per email sent, watch diminishing returns. Ecommerce 2-4/wk engaged; newsletter 1-3/wk; SaaS 1-2/mo.

11. DELIVERABILITY TRIAGE

Authentication (all required): SPF (end -all, 10-lookup limit) · DKIM (2048-bit, rotate yearly, aligned) · DMARC (p=none → quarantine → reject; Outlook no longer accepts p=none for 5K+/day, and rejects non-compliant bulk with a 550). BIMI/VMC is the under-used Gmail trust lever (worth it for top-tier senders with enforcement + trademark + VMC).

Reputation: domain > IP for Gmail (120-day memory). Dedicated IP only at 1M+/month. Separate marketing and transactional subdomains at 40K+/month.

Diagnosis path (when placement drops): symptom → check auth → blocklists → reputation → bounce logs → sending patterns → content → test/validate → fix root cause → monitor recovery (2-4 weeks, Gmail up to 120 days).

Thresholds with actions:

• Complaint rate ≥0.1%: pause broad sends, restrict to clicked-30d, inspect acquisition source and expectation mismatch, confirm unsubscribe visibility.
• Engagement a primary signal now: a chronically unengaged list out-loses a smaller engaged one. Auto-sunset the dead weight.
• "Low bounce" ≠ "safe to send": consent and engagement signals can suspend an account at 0.1% bounce.

AI-era deliverability:

• Gmail Gemini summarises mail and re-ranks Promotions by relevance; AI summaries decide the preview from the first ~150-200 characters of live text, overriding your preheader. Front-load purpose in real text, use semantic HTML, never image-only, consider Gmail JSON-LD annotations.
• Raw un-personalised AI text is filtered harder (Google's gen-AI spam filter). Personalisation tokens are now a deliverability requirement.
• "AI does not hide cracks, it exposes them." Clean data + auth are the prerequisite to agentic sending, not an afterthought.

Autonomous-send guardrails: human-in-the-loop for any blast; hard volume caps on AI-triggered flows; mandatory engagement-tier targeting even when an agent composes; the ESP should surface reputation/spam-rate to the agent before it sends. (See §0.)

Warm-up: ramp engaged-first, staggered over your normal window (e.g. 20→80/day over two weeks for new sender identity, or scale a domain 300→500→…→10K/day over ~14 days). Continue warming alongside live sends. Switching ESPs: verify the list, warm by sending most-engaged first in chunks, re-opt-in 6-month-dormant contacts.

12. TESTING & OPTIMISATION

• Highest-value tests: sender name (compounds), CTA format, template structure. ~1 in 7 tests yields a significant winner; use 95% confidence. Test flows over campaigns (improvements compound).
• Testing AI-assisted email: guard against homogenisation (AI variants converging on the same safe phrasing). Run an explicit anti-slop test, measured on reply rate and Primary-tab placement, not opens.

13. COMPLIANCE GATES

Decision gate before any send: (1) type (transactional / lifecycle / marketing / newsletter / cold)? (2) recipient region? (3) consent basis? (4) unsubscribe + physical address present? (5) suppressions applied? (6) content materially accurate? If any is unclear, refuse or ask, do not proceed to approval.

| Regulation | Consent | Key rules | Penalty | |---|---|---|---| | CAN-SPAM (US) | No | accurate headers, physical address, honour opt-out ≤10d | ~$51,744/email (as of 2026) | | GDPR (EU) | Yes | erasure 30d, consent records | up to 4% turnover / €20M | | CASL (Canada) | Yes | implied consent 2yr after purchase, express = indefinite | up to $10M CAD | | Spam Act (AU) | Yes | consent + sender ID + unsubscribe ≤5 biz days | up to $2.22M AUD/day |

One-click unsubscribe (RFC 8058) required for 5K+/day to Gmail/Yahoo/Microsoft; honour within 48h. AI does not transfer liability: you are accountable for an agent's sends; an AI can run a pre-send compliance pass but never trust it to preserve the unsubscribe/footer when it edits a template. Cold email: B2B legal in US/UK without consent, consent required in Canada/Australia.

14. COLD EMAIL

• Infrastructure: never send from your primary domain. Separate domains, warm 2-4 weeks, 10-30/inbox/day, a dedicated cold tool (not your marketing ESP). Keep it legally and infrastructurally separate from marketing.
• Writing: 50-125 words. Personalised opening → observation/problem → value → soft, interest-based CTA (2-3x the replies of a meeting ask). It must still read like one human emailing another, AI-personalised or not.
• Follow-up: 4 emails over 2-3 weeks, each adding new value; breakup gets 2-3x the reply rate.
• AI in outbound: AI prospecting/personalisation (2-3x reply vs templates, produced faster) and autonomous reply handling, with the same domain/suppression/consent guardrails. Founder-led 1:1 from a real inbox still beats cold-blast on B2B reply + deliverability.

15. PLATFORM SELECTION

Selection factors: ecommerce depth · event/data model · **AI + programmatic interface (can an agent drive it via MCP/app, or dashboard-only;

…

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: CosmoBlk
• Source: CosmoBlk/email-marketing-bible
• License: MIT
• Homepage: https://emailmarketingskill.com

Install and usage instructions live in the source repository linked above.