FRDTLAB · Field Note · How it works
Marketing through the system, not on top of it.
If you run social, SEO, and paid for a business by yourself, you already know how to build a campaign structure. An AI operating layer is the same discipline pointed at your own work. Here is the whole thing in one picture and six stages, in words you already use.
The one idea. The model is not the system. It's the router. Everything that makes the output sound like you, use your data, and land where you actually publish lives in the layers around it. Most people set up the router and skip the layers, then wonder why every session starts from zero.
The six stages, translated
You already do every one of these in PPC and SEO.
| Stage | What it actually is | You already do this in PPC / SEO | What breaks without it |
|---|---|---|---|
| 1 · Brief | Standing context that loads before you type: who the brand is, what's decided, what's in flight, where files live. | Your account structure and naming convention. Nobody re-explains the account every morning; it's in the structure. | Every session starts cold. You re-paste the brand guide. Output drifts. |
| 2 · Route | Rules that pick the voice, the guardrails, and the review steps for this kind of task. Encoded once, applied every time. | Campaign settings and negative keyword lists. Match types. The stuff that stops the wrong thing from serving. | Client tone leaks between brands. Health or legal claims slip through. You become the QA step. |
| 3 · Sources | Live pipes into your real data: analytics, Search Console, ad platforms, the CMS, your own files. Read where they sit; don't copy. | Conversion tracking and data feeds. Merchant Center reads your catalog; you don't retype products into ads. | The model guesses. Numbers in reports can't be traced to a source. |
| 4 · Work | The actual task: draft, analyze, compare. Small jobs run inline; big ones fan out (read 40 pages in parallel) or run as a pipeline (draft, review, verify). | Bulk edits and scripts. You don't hand-edit 400 ad groups; you run the sheet. | Everything is one long chat. Nothing is repeatable. |
| 5 · Publish | Output lands where it's used: the site, the scheduler, the report, the task board. The chat is the preview, not the destination. | Push to live. A campaign in a spreadsheet earns nothing until it's in the account. | Deliverables die in chat tabs. You copy and paste all day. |
| 6 · Remember | Decisions, learnings, and open loops get written down in the same turn, and scheduled runs keep working when you're not there. | Change history, automated rules, scheduled reports. The account keeps running while you're in a meeting. | The system never compounds. Month two is as slow as month one. |
Why the drive is the bus
One copy. Many readers.
Every tool you use has its own memory: the chat app remembers your account, the desktop app remembers this machine, a session remembers itself until it closes. None of those can see each other. A shared drive is the one place all of them, plus your automation tool, plus a scheduled run at 7am, plus a human on your team, can read and write the same file.
You already trust this pattern. Your product feed is the single source of truth that Merchant Center, the ad account, and the site all read from. Nobody maintains three catalogs. The operating layer just applies that rule to how the business works: brand rules, decisions, briefs, deliverables. One copy, many readers.
If you're a one-person shop
Set up these three things first.
A registry file
One document that lists every brand or client you touch: a short code, where its files live, what the current gate is, who owns it. Any tool that reads it starts informed.
A voice kernel per brand
Not a 40-page brand book. A one-page rule set: register, banned words, claims you can't make, the CTA pattern. It's what keeps a local clinic's copy from sounding like a parts retailer's.
A naming rule
Files, tasks, and chats carry the brand code and the date in the name. Naming is routing. A file called "final-v3.docx" belongs to nobody; a file called ACME_Report_2026-09.docx files itself.
Do those three and the rest (data pipes, scheduled runs, task-board mirrors) has somewhere to attach. Skip them and you're building automation on sand.
What changes about your job
You hold judgment. The system holds state.
- You stop being the memory. The registry and the drive hold state. You hold judgment.
- You stop being the QA step. Guardrails are encoded in the route stage; the system checks health claims, banned words, and source citations before you see the draft.
- You review at the gate, not the keyboard. Agents draft and stage; nothing client-facing ships until a human flips it. That single rule is what lets the rest run unattended.
- Your reporting gets honest by default. Every number traces to a pipe. Gaps get flagged as "awaiting data" instead of being smoothed over.
Three things people get wrong.
- "Better prompts." A prompt is one request. The layer is what makes the thousandth request as good as the first without re-explaining. Prompts are stage 4; most of the payoff is in stages 1, 2, and 6.
- "One tool to rule them all." You'll use several. The question isn't which one; it's whether they all read the same drive. If they do, switching costs nothing.
- "Automate first." Automation runs on top of structure. A scheduled run can only act on brands whose brief, voice, and home are filled in. Empty slot, skipped work.
Bottom line. You are not learning a new discipline. You are applying account structure, negatives, feeds, and change history to your own operation. The model does the typing. The layers do the remembering. That's marketing through the system.
Educational, not a tool recommendation. More operating models live in Frameworks.
From field note to build
Want the layer built with you?
The registry, the voice kernels, the pipes, and the gate, installed alongside your team.