Field Notes: Running Marketing for 11 Products, Alone

01The constraint is the strategy

Eleven products, one person, effectively zero ad budget. That constraint forces a decision most marketing teams never have to make explicitly: every hour of human attention has to go where judgment matters, and everything else has to run without me.

So I stopped thinking in campaigns and started thinking in systems. A campaign is something you do; a system is something that keeps doing it while you're quoting a painting job or sitting in a discovery call. The rest of these notes describe the three systems that do the heavy lifting, and the one rule that keeps them honest.

02The division of labour

The split isn't "AI does marketing." The split is a line between production and judgment, drawn deliberately:

Agents produce

First drafts of outreach and follow-up variants
Data collection & report assembly for content
Funnel health checks on a 3-hour heartbeat
Lead-list enrichment and pipeline hygiene
Draft replies queued for review, never sent

I decide

Positioning, pricing, and what gets built
Which channel gets the next push, and why
Every word a customer actually receives
When to kill a channel that isn't working
What "working" even means — the metric

The output side of that table is where the leverage is. The right side is why it's still marketing and not spam.

03Three systems, real numbers

i. Content that starts as data, not opinions

For LinkRescue (a broken-affiliate-link monitor), the content engine is a monthly Link Rot Index: agents crawl a fixed panel of content sites, check every outbound affiliate link, and assemble the raw findings. I turn the findings into the editorial angle, the post, and the programmatic SEO pages.

6,550

links checked across 50 sites and 683 pages in the June issue

5.8%

of links broken outright — the headline stat the post is built on

58%

of attribution issues were silently lost tracking parameters — the angle nobody else had

The point: original data is the one content moat a one-person operation can afford. Agents make the data collection free; the insight still has to be found by a human reading the results.

ii. A funnel where every stage reports its numbers

AI Commerce Radar (an AI-visibility audit for Etsy/Shopify sellers) launched as a complete measured funnel: free audit → scored, shareable report with social cards → email capture → paid subscription. Before driving any traffic, I set a 14-day validation gate with pre-committed thresholds — the funnel either earns more investment with data or gets parked without sentiment.

An agent heartbeat checks the funnel's health every three hours and pushes a card to my phone when a stage breaks or moves. I don't check dashboards; the dashboard checks in with me.

iii. Outbound where personalization is manufactured

For PaintPulse (SMS photo-updates for painting contractors), the classic outbound trade-off is volume versus personalization. The system breaks the trade-off: for each prospect, an agent seeds a live demo pre-loaded with that company's own branding before the first email exists.

166

painting contractors researched into the lead database

personalized demo sites seeded before first contact

1:1

every email references a demo that already exists for that prospect

The email itself is short, because the demo does the talking. Cold outreach converts on proof-of-effort, and agents make effort cheap to prove.

04An annotated prompt

Prompting is the visible 10% of AI skill, but it's the part people ask about, so here's a real pattern — the outreach drafter, condensed. The annotations are the actual craft:

You draft cold emails for {product} to {prospect}. ①
Return 3 variants as JSON matching the provided schema. ②

Context you may use: {prospect_research}, {demo_url}, {icp_notes} ③

Rules:
- Max 90 words. The demo link carries the pitch. ④
- Reference one specific, verifiable fact about the prospect.
If research contains none, say so and stop. Do not invent one. ⑤
- No "I hope this finds you well." No exclamation marks.
Write like a contractor texting another contractor. ⑥
- Output is a DRAFT for human review. Never imply it was sent. ⑦

①Role + variables up front. The template is reused across products; only the variables change. Prompts are infrastructure, not one-offs.
②Structured output. JSON against a schema means drafts flow into the review queue programmatically — no copy-paste step to break.
③Explicit context boundary. The model may only use research it was handed. This is the #1 defence against confident nonsense.
④Constraint carries strategy. The 90-word cap isn't style — it encodes the insight that the personalized demo, not the email, closes.
⑤A defined failure mode. The most important line in the prompt. An agent that can say "not enough information" is worth ten that can't.
⑥Voice by example, not adjective. "Like a contractor texting another contractor" outperforms "friendly and professional" every time.
⑦The guardrail, restated in-prompt. Belt and suspenders: the pipeline already can't send, and the prompt knows it too.

This pattern isn't theoretical — a version of it runs live on this site. Paste a cold email into the grader and watch it score against the same rubric.

05The rule that keeps it honest

Nothing outward-facing ships without a human decision. Not an email, not a post, not a deploy. Agents draft; a person sends.

This isn't caution theatre — it's the operating rule that makes the whole system defensible. Every automated step leaves an audit trail; every customer-facing word passed through human judgment. When people ask what "responsible AI use" looks like in marketing, I think it looks like this: maximum leverage on production, zero delegation of accountability.

Stack, for the curious: Claude API & agents · HubSpot CRM · Google Analytics 4 · Next.js + Vercel · Stripe · a Postgres database and a pile of cron jobs.

Running marketing for eleven products, alone.