The campaign brief was one paragraph. “Write a nurture email for mid-funnel prospects — something warm, conversational, enterprise-ready.” The AI returned five variants in forty seconds. Clean, confident, well-punctuated. The campaign manager liked the tone, made two small word-swaps, and sent them to a list of 22,000 contacts.
Three days later: two variants had claimed a capability the product did not yet offer. One addressed the reader as a startup founder — correct for an earlier ICP, wrong for this segment of VP Engineering at mid-market companies. None had a measurable call to action. Performance was flat, but the compliance problem was the expensive one: a customer had forwarded the email to their legal team.
This is the signature failure of AI-assisted marketing. Not a hallucination — the claim was almost true, for a forthcoming release. Not a style failure — the copy was polished. The failure was strategic: on-brand-looking copy that was off-strategy, off-segment, and made a claim nobody had authorized. It looked done, so nobody caught it until it was out.
The software-native version of this method calls this fast waste. In marketing, fast waste has a specific texture: volume at velocity, with the errors invisible until the emails land, the ads run, or the legal flag arrives.
The four failures, in marketing
The ADD Beyond Code framework names four failures that recur when AI produces generative work without structure. Each has a marketing form.
Fast waste, branded. Given a vague brief, the AI produces copy that reads well but is aimed at the wrong segment, makes an unsupported claim, or contradicts the current campaign strategy. Because the output is polished, it survives a casual read and ships.
Context rot. The brand voice guide was last updated when the company had thirty employees. The ICP document predates the pivot to mid-market. The approved-claims list lives in a shared doc that legal updated in Q2 and marketing has not re-read. Every new generation session re-derives these constraints from scratch, inconsistently.
Trust-by-inspection breaks down. A copy variant that sounds right is not evidence that it is right. It sounds like your brand because the AI has seen your brand. It will also sound like your brand while making a claim legal has never approved, addressing a reader outside your segment, and omitting the CTA your campaign requires.
Verification ceiling. When AI produces fifty variants overnight, the bottleneck is no longer generation — it is review. If the review process is “the campaign manager reads them,” throughput is capped at what one person can carefully evaluate. Volume beyond that is not marketing output. It is unreviewed risk at scale.
The loop, translated for marketing
ADD’s eight steps map onto a campaign production flow. The concrete artifacts are where the method lives or dies.
0 · Ground: load the real context
Before writing a word of copy, load the context the AI will stand on: the current brand voice guide, active ICP and segment definitions, approved claims with evidence sources, campaign objectives and target metric, and recent learnings from past sends.
Ground is the antidote to context rot. The strategy and constraints do not live in the campaign manager’s head or in the chat thread — they live in one short document the brief reads first. When a new campaign begins, that document re-orients the session instead of having the AI re-guess what your brand sounds like.
1 · Specify: the brief as a spec
Most creative briefs are directional. They describe the feeling, the audience, the rough goal. An ADD brief is a spec: it states what the copy must do, what it must not do, and names a reason for each prohibition.
Named refusal reasons are the move that makes this work — not style guidelines but machine-checkable rejection criteria, each paired with a code the AI invokes when it cannot comply.
# CAMPAIGN BRIEF — Q3 Mid-Market Nurture: "Time-to-Value"# Version: 1.0 · Status: DRAFT (awaiting brand+legal freeze)
Segment: VP Engineering, companies 200–2,000 employees, currently evaluating (mid-funnel)Goal: Drive demo bookings; target metric = email click-to-demo conversion rateAfter-state: The reader books a demo call feeling that we understand the cost of slow deployments
Must: - Communicate that customers reduce deployment cycle time - Address reader as a technical decision-maker, not a business buyer - Include one clear CTA linking to /demo with UTM parameters - Stay under 180 words (mobile readability) - Use plain, declarative sentences (see brand voice guide §3)
Must NOT: - CLAIM-UNSUBSTANTIATED: Do not cite specific performance percentages (e.g., "50% faster") unless the claim appears in approved-claims.md - OFF-BRAND-VOICE: Do not use superlatives ("best," "leading," "revolutionary") or exclamation marks — brand voice is calm and precise - WRONG-SEGMENT: Do not address startup founders, C-suite executives, or non-technical readers; segment is VP Eng at mid-market - NO-CTA: Every variant must include a clear next step with a link; copy that ends without a CTA is rejected - LEGAL-UNCLEARED: Do not mention the roadmap feature "automated rollbacks" (not GA; flagged by legal 2025-11-03)
Assumptions — lowest-confidence first: ⚠ "reduce deployment cycle time" is approved as a directional claim, but verify against approved-claims.md before freeze; if not present, remove or escalate to legal before generation begins.The assumptions section matters. Flag the claim you are least certain about and confirm it with brand or legal before generating — one sentence of human attention that eliminates an entire class of compliance failure.
2 · Scenarios: the model case, the edge, the failure
Three scenarios make the brief concrete before a word is written.
The model case: A VP Eng at a 500-person SaaS company reads the email on a phone, three weeks into evaluating two vendors. The email is short, addresses deployment pain directly, and links to the demo page. They click.
The edge case: A skeptical reader who has already seen three competitor emails this week. The angle must give them something concrete — a specific problem named, a real trade-off — not a list of benefits. Benefit-soup triggers deletion.
The failure case: A variant claims “deploy 3× faster” without an approved source. It passes a casual read, ships, and a prospect asks for the study behind the claim. There is no study. CLAIM-UNSUBSTANTIATED.
Scenarios surface the judgment calls — the edge reader, the hostile reader — that a checklist cannot fully anticipate.
3 · Contract: the frozen brief, the one human gate
The brief becomes a frozen contract once brand and legal sign off. That sign-off — approving the segment definition, the allowed claims with evidence sources on file, the refusal codes, and the voice constraints — is the single human gate before generation begins. After freeze, the brief does not change mid-campaign. A needed change is a change request: return to legal, re-freeze at a new version.
Constrain the what (audience, message, claims, CTA) and leave the how (angle, tone, structure, opening line) wide open. How the email opens, what analogy it reaches for, whether it leads with a question or a statement — all unconstrained and disposable. What does not move is the target, the claim set, and the CTA.
4 · Acceptance checks: the “red tests” for copy
Before generating, write the checklist the output must pass. These are domain-specific acceptance criteria — the marketing equivalent of a failing test suite.
# ACCEPTANCE CHECKLIST — Q3 Mid-Market Nurture# All items must pass before a variant is eligible for send
Voice & brand [ ] No superlatives or exclamation marks (OFF-BRAND-VOICE) [ ] Plain, declarative sentences; no passive constructions as lead [ ] Reading level appropriate for a technical VP (not a business buyer)
Claims [ ] Every factual claim cross-referenced against approved-claims.md [ ] No performance percentages unless sourced (CLAIM-UNSUBSTANTIATED) [ ] "automated rollbacks" not mentioned (LEGAL-UNCLEARED)
Segment fit [ ] Addresses VP Engineering, mid-market; no startup or C-suite framing (WRONG-SEGMENT) [ ] Deployment cycle time is the primary pain; product features are secondary
CTA [ ] One clear CTA present, linked to /demo with correct UTM parameters (NO-CTA) [ ] CTA is the last element; no content after the link
Format [ ] Under 180 words (body only, not subject line) [ ] Subject line under 55 characters (mobile preview)
Compliance refute-read [ ] A second reader has specifically argued why this variant should NOT send [ ] Any raised concern addressed or escalated; no silent overrideThe compliance refute-read is the adversarial step. Before any variant ships, a named reviewer makes the case for why this specific variant should not go out — not “does anyone have concerns?” which invites silence. If they cannot find a reason, the variant passes. No silent skips.
5 · Produce: generation inside fixed walls
Now the AI generates variants against the frozen brief and the checklist. The how is unconstrained: the opening line, the structural choice, the specific analogy. A variant that cannot pass a check invokes a refusal code and explains why — more useful than one that silently violates a constraint and looks fine.
6 · Verify by evidence: the A/B test, not the read-through
A campaign manager reading five variants and choosing the one they like is not verification — it is preference, and it reproduces the exact failure ADD is designed to eliminate: plausible-looking output trusted by inspection.
Verification in marketing is the A/B test against a control. Ship the top two variants to a split of the list. Measure click-to-demo conversion rate — the target metric named in the brief — against a recent comparable send. The verdict is evidence, not consensus.
Secondary verification: the compliance refute-read. Someone specifically argues against each variant before send — the five-minute step that catches the problem before 22,000 contacts receive it. “The team likes it” and “it reads well” are signals that the output is competent. They are not evidence that it is correct.
7 · Observe and fold: performance becomes the next brief
After the campaign runs, the performance data becomes structured input to the next brief: which segment responded best, which angle drove conversion, which subject line pattern underperformed, which claim generated follow-up questions that signaled it needed evidence.
This folds back into a living brand playbook — a versioned document that grows with each campaign, tagging confirmed patterns and deprecating failed ones. Not a post-mortem nobody re-reads. A document the next brief reads first. Context rot dies here: the next campaign manager inherits confirmed knowledge rather than re-deriving what works from scratch.
What doesn’t transfer
ADD’s marketing translation is genuine, but three limits are worth naming honestly.
Brand taste has irreducible human judgment. The brief can specify calm and precise, but it cannot specify distinguished. The difference between copy that avoids violations and copy that genuinely resonates with a skeptical VP Eng is not checkable. A brand that over-templates its generation will produce compliant work that is, across the portfolio, undistinctive. The checklist tells you what must not fail; it does not tell you what makes something good.
Not everything is fast A/B-testable. A brand launch campaign, a founder letter to customers, a thought-leadership piece — these run once and the signal is slow. Treating every campaign as a rapid conversion experiment flattens the register of things marketing does.
Over-specifying the brief kills the angle. If the contract constrains not just the what but also the opening line, the analogy, the structure — it has crossed from spec into transcript. The brief is too loose if it produces WRONG-SEGMENT variants; it is too tight if every variant opens with the same construction. An ADD brief takes more time to write than a vague prompt, and finding that calibration is the actual skill.
Next in the series
The same pattern applies wherever AI produces generative work at volume. In sales, the problem is AI-drafted outreach that is personalized-looking but off-ICP, and the equivalent of the claims-substantiation check is a pipeline-outcome audit.
Next: ADD for Sales: Sequences as Specs, ICP as the Frozen Contract.