Email Marketing Bible

Email Marketing Bible, Skill Reference

By George Hartley, co-founder of Nitrosend.

v2.7, 8 Sep 2026. Distilled from the EMB (19 chapters, 908 sources, https://nitrosend.com/email-marketing-bible), from running SmartrMail (~12K customers, 6B emails, sold 2022) and three months running Nitrosend through agents. Part A is the operating manual, Part B the reference. Figures are mid-2026; verify anything volatile (inbox rules, ESP features, pricing, model names) before acting.

---

PART A: OPERATING MANUAL

0. AGENT OPERATING RULES

Every segment, draft, campaign, flow or staged send on a real ESP is live. Hard gates, never skip:

• No 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.
• Preview before asking; show the packet before any send: preview URL, audience size, exclusions/suppressions applied, subject, preview text, send time, from-name + reply-to, unsubscribe present, compliance risk.
• Block the send if authentication is missing, unsubscribe or physical address is absent, complaint rate is at or above 0.1%, consent basis is unclear, or the audience includes suppressed, bounced or complained contacts.
• Never probe unknown mutating endpoints on a live audience. /send, /dispatch, /trigger, /fire, /publish paths can dispatch immediately; if the approve-scheduled path is unclear, ask the human to click it. Test on sandboxes or cloned campaigns with seed lists.
• Separate the modes. Transactional, marketing, lifecycle and cold outbound have different rules, domains and consent bases. Never mix them.
• Log every autonomous action (segment changed, flow edited, campaign created, send staged) so the human can audit it.

1. TASK ROUTER

2. AI EMAIL AUTOMATION (the operating model)

The marketer moved from operator to director: brief the agent, govern it, own the send button. Most major ESPs now ship a human-gated prompt-to-campaign agent, an MCP server or a Claude/ChatGPT app (§15); advise on the surface the user runs.

The loop: read state → reason → act → verify. Read the account first (lists, flows, recent campaigns, deliverability, suppressions), act on one thing, verify it. Opening prompt: "audit my account and tell me what is missing".

Automate: send-time optimisation, subject-line variants + A/B, cart/browse triggers, post-purchase cross-sell, first-draft copy. Keep human: brand voice, strategy (segment priority, flow order), creative direction, domain and deliverability, the final send.

Autonomy dial. Ask mode by default; widen only on narrow, reversible, low-brand-risk tasks, with an undo; read before write access. Supervised autonomy is the production stance. Where "AI optimisation" means bandits reallocating live traffic, measure with holdouts (never last-touch credit) and do-not-optimise constraints (margin, fatigue, complaints, brand safety).

Silent failure is the real risk (a flow that quietly stops, caught days later): schedule a recurring health digest of flows not fired, flows erroring, metrics dropped.

2b. FIELD NOTES: RUNNING AN ESP FROM AN AGENT (JUN-SEP 2026)

Three months running Nitrosend's own sending through agents; each rule cost a real mistake. First five are Nitrosend mechanics (check your ESP's equivalent); the rest hold anywhere.

• Optimistic-concurrency version (if_version) on every write; on conflict, re-read and retry with the fresh version, never guess.
• Re-assert brand or account before every write batch after idle; MCP context resets silently to the default brand while reporting a deliberate selection.
• Silent-parameter APIs default to send-to-all: pre-flight assert audience id and count, never probe a mutating endpoint on a live audience (one unknown body key mailed 1,003 contacts).
• Liquid merge defaults go unquoted inside href; inner quotes close the attribute and break the link.
• Animated WebP rather than GIF for heroes; then fetch the served URL and confirm it still animates (CDN variants can flatten to frame one).
• Set text and button text colours explicitly on every design; theme defaults drift (grey headlines, dark text on a coloured button).
• Decode tracking-wrapped CTA URLs before approving; the wrapper hides the target.
• Never backfill or re-dispatch failed sends without a human order; late sends look worse than none.
• Drafts by default; the literal "send it" in chat is the only thing that fires a blast.
• Every email gets a hero, a live-text headline and one button; secondary content gets inline links.
• Quote tiles come from HTML in headless Chrome, never an image model (garbled type, invented names).
• Migration opt-out state comes from the old ESP's API, never a list CSV; exports drop unsubscribes.

3. PRE-SEND CHECKLIST

Confirm every line, surface it, wait for "send it".

• Audience: size and segment logic verified against actual counts (AI segments run over-broad)
• Suppressions: unsubscribed, bounced, complained, globally suppressed, frequency-capped, open support issue
• Authentication: SPF, DKIM, DMARC aligned, p=quarantine or stronger (Outlook requires all three at 5K+/day)
• One-click unsubscribe (RFC 8058) + physical address present
• Copy: §4 pass, one CTA, subject ≤45 chars, preview text adds information
• Design: single column ≤600px, dark-mode safe, alt text, live-text headline, explicit text and button colours, images <200KB each and <800KB total, cross-client preview, spam score, hero animates at the served URL
• Links: wrapped CTAs decoded, no placeholder URLs, merge defaults render inside href
• Sender: correct from-name + monitored reply-to; brand and account re-asserted; send time set; consent basis valid for this audience and content
• Non-email: US SMS 10DLC brand + campaign registered; WhatsApp opt-in for the category + approved template; quiet hours per recipient local time (SMS 8am-9pm)
• Kill switch: batched or throttled send with a working pause and rollback plan
• Test send reviewed in a real inbox with real merge data
• Personalisation confidence, inventory and pricing freshness checked; kill plan named
• Human approval captured

4. ANTI-SLOP COPY PROTOCOL

Raw LLM copy is a deliverability liability, not only a quality one: Google filters high-AI-similarity text harder.

• The deepest tell is the absence of stakes. Put one genuine, defensible opinion in every email. Ask the draft where it is too safe.
• Burstiness. Alternate long and short sentences; a 3-5 word line 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", rule-of-three padding, copula avoidance ("serves as" for "is"), em dashes.
• Specificity is the cheapest humaniser. Real numbers, names and dates. Pull one real metric from the brand's own data into every email.
• Workflow: human strategy → AI draft → human edit. High-personality formats (founder letter, welcome): rough human notes first, AI tightens. Never AI-first.

5. AI EMAIL DESIGN PROTOCOL

AI defaults to competent and generic; force it off its defaults.

• Two readers: the human and the summariser. Gmail's Gemini and Apple Intelligence summarise from the opening live text (rollout tiered). Front-load the offer in real text, semantic headings, never image-only; live text also wins accessibility and dark mode.
• Context beats prompt. Feed brand kit, design tokens, a tested module library and a rules file before iterating on wording.
• Safe substrate. Emit MJML, React Email or Maizzle (compile to inbox-safe HTML), never raw HTML from a prompt.
• Anti-slop design rules: own one colour (30-60% of the surface); restraint over decoration; real photography, never AI stock; bold live-text headlines; one message, real negative space. Ban the purple-to-blue gradient and the beige wash.
• Compliant by default: single column ≤600px, 44px tap targets, role="presentation" tables, dark-mode-safe colours (~#121212, never pure #000 backgrounds or #fff logos), alt text everywhere, explicit text and button colours.

Direct the agent: Discover, Define, Deliver. Adapted for email from Anshu Chimala, "How to turn your AI into a world-class designer" (Lenny's Newsletter, 1 Sep 2026, https://www.lennysnewsletter.com/p/how-to-turn-your-ai-into-a-world) via the design-director skill (command, counts and brief format are the skill's). LLMs predict the median; divergence has to come from outside the model.

• Seed strings. The agent generates a random string in a shell (openssl rand -base64 48), derives palette, layout and type from its patterns, never reveals it; new string per direction.
• Broad before deep. Ask for 12-20 directions as one-liners, "go broad, not deep". The human picks from text before any image or code exists. Reject anything guessable from the category alone.
• Ambitious briefs. One sentence naming a real reference (Graza's chartreuse drench, Aesop's restraint) plus two anti-references.
• The critic loop. Screenshot the rendered test send and hand it to a separate, stronger model in a fresh context (no code, history or earlier critiques). It names the aesthetic, imagines how a top studio would execute it, lists the biggest gaps and scores /10. Fix, re-screenshot, re-critique with the same prompt (target score kept out of it) until the critic scores 9/10, capped at four rounds. The critic is ~10% of output tokens and most of the taste.
• Chain models. Code model for structure, image model for stills, video model for a looping hero or state transition. As of Sep 2026 (verify): Claude Fable 5.1 as critic; Claude Opus 5 or Sonnet 5 (Claude Code) or GPT-6 Astra (Codex CLI) as implementer; gpt-image-2 for stills; Gemini Omni 1.1 for video.
• Deliver by subtraction. Cut glows, gradients, decorative containers and labels that repeat the visual, then a light anti-slop pass on copy (§4) and visuals (reflex fonts, centred hero + three cards, purple on dark).
• Keep failed prompts; retest on the next model generation.
• Who to follow (Chapter 18, 49 practitioners, five added in v2.7): Anshu Chimala @anshuc, Karri Saarinen @karrisaarinen, Ryo Lu @ryolu_, Jenny Wen @jenny_wen, Lee Munroe @leemunroe.

---

PART B: REFERENCE

6. FUNDAMENTALS & METRICS

Owned media at ~$36 per $1; 5K engaged beats 50K messy; flows before campaigns.

Open rate is noise. MPP pre-loads pixels and Gmail/Apple summaries auto-open mail (opens inflate while CTR falls). Judge on clicks, replies, conversions and revenue per recipient; label open-only reads low-confidence; never compare opens across ESPs.

Lists vs tags vs segments: one master list; tags are facts; segments are dynamic rules. Minimum segments: new (30d), engaged (clicked 60d), customer vs non-customer, lapsed (90d+).

7. CORE FLOW RECIPES (the revenue engine)

Flows out-earn campaigns ~30x per recipient. Build in this order (you specify trigger → wait → condition → send; the agent scaffolds; you review):

Welcome → Abandoned cart → Browse abandonment → Post-purchase → Win-back → Cross-sell → VIP → Sunset → Birthday → Replenishment → Back-in-stock → Price drop.

• Welcome (4-6): promise + reply ask + one segmenting question → brand story → social proof → best content by answer → soft sell → expectations. 51-55% opens.
• Abandoned cart (3): reminder, no discount (1-4h) → objections: reviews, shipping, guarantee (24h) → small incentive if margins allow, first-timers only (48h). ~17% recovery.
• Post-purchase: confirm → shipping → satisfaction check → review → cross-sell → replenishment.
• Win-back (60-90d inactive): "we miss you" → value offer → breakup (highest reply) → confirm + resubscribe.
• BFCM: build list (Sep-Oct) → warm volume (Oct-early Nov) → tease (2-3 wk out) → daily sends, engaged first → post-BFCM thanks, cross-sell, shipping deadline.
• Consistency beats perfection: a 20-minute weekly (Liz Wilcox) or 2-3 short sends beat one polished monthly (Ian Brodie).

8. COPYWRITING REFERENCE

• Subject lines decide the open: under ~25 chars opens highest; lowercase casual can beat title-case (~14%); first-person CTA beats second-person.
• Body: inverted pyramid, short paragraphs, write then cut 30%. 3:1 value-to-promo.
• Frameworks: AIDA (promo) · PAS (cold/B2B) · BAB (case studies) · Soap Opera Sequence (narrative) · 1-3-1 newsletter (one story, three items, one CTA).
• CTAs: buttons beat text links (+27%); one CTA beats several (+42%); above the fold and below the main content.

9. SEGMENTATION & LIST BUILDING

• Personalisation hierarchy (high → low): behavioural → lifecycle stage → dynamic blocks → send time → location → name. Above all: agent-generated 1:1 content from real behavioural data (clean data first; draft-and-approve).
• Segments from natural language: let the agent build the rules, then verify against actual counts before sending. A segment that jumps 10x between runs is a bug until proven otherwise.
• Engagement-based sending (highest-impact lever): clicked 30d → every send; 60d → 75%; 90d → best only; 90-180d → re-engagement only; 180d+ → sunset. Opens +15-30%, complaints -20-40%, revenue holds or rises.
• List building: lead magnets (templates convert best) · content upgrades (5-10x sidebar forms) · forms beat links (+20-50%). Popups 3-5% (top decile ~9%); exit-intent 4-7%; two-step beats one-step. Double opt-in for lead magnets, single for purchasers.
• Hygiene: lists decay 22-30%/yr. Sunset: reduce frequency → 2-3 re-engagement emails → suppress. Trap prevention: double opt-in, real-time validation, engagement-based sending.