Define the offer before you touch a single tool
Every AI marketing failure Matteo has audited traces back to one thing: an offer nobody defined clearly first. Write down who you sell to, what you sell, what makes it worth buying, and what a new customer is worth. Then run the transform habit on every request after this one:
Then define the offer itself:
The reasoning matters more than the template. An AI system with no clear offer writes generic ad copy, reaches for the wrong audience signals, and optimizes toward the wrong number, because it has nothing specific to optimize toward. A $40 skincare product selling on a first-purchase margin needs a different cost ceiling than a SaaS free trial with a 90-day payback window. Get the offer wrong here and every step after this one inherits the mistake, no amount of clever prompting downstream fixes an offer nobody defined.
"We sell a $40 skincare serum to women 28 to 45 who already use retinol and want fewer breakouts without the irritation. Our edge is a dermatologist-formulated blend at a drugstore price. A new customer is worth roughly $140 across their first year once repeat purchase is factored in, so a first-purchase cost per sale up to $45 to $50 still pays back." Specific enough that an AI, or a media buyer, can set a real cost ceiling directly from it.
Save it before moving on, so step two has something real to read:
Give your AI a memory
Most people start every conversation from zero. A folder of plain text notes, kept in a free app like Obsidian, becomes the memory your AI reads before every task: the offer, the proof, the brand rules, what worked and what failed. The full five-folder structure sits further down this page. Once it exists, point your AI at it:
Every session after this one starts already knowing the business.
The five-folder structure works because it separates what changes fast from what changes slow. People and Projects update by the week. Swipe and Results are append-only logs that only grow, once a campaign ends, write down what happened and leave the entry alone rather than editing it later.
Every note gets a date and a reason. "(2026-06-01) Killed the testimonial-lead angle after 11 days, CTR held at 0.6 percent, frequency capped it before a thousand impressions" tells a future session what happened and why, well beyond just "testimonial ad didn't work." Never write a secret, an API key, or a password into a notes file, keep those in a separate credentials vault.
Teach it to write back, not just read:
Turn prompts into skills
A prompt works once. A skill works every time, the same way, for as long as it stays installed. Here is the real ad-creative skill, trimmed:
This one and the prompt-engineer AI agent from step one are the two free AI agents waiting in your download below. Dozens more like them live inside Full Stack AI Marketing.
Write the description field as a trigger phrase that names exactly when it should fire. "Use when writing ad copy, planning creative, or deciding what to test next" tells the AI precisely which requests should load this skill without you naming it by hand. A vague description like "helps with marketing" fires unreliably, and an unreliable skill goes unused.
Once a skill grows past a page, split it. Keep the main file as a short entry point, the core principles and the trigger conditions, then move the deep material, hook libraries, testing calculators, brief templates, into linked reference files the main file points to. A skill anyone can scan in ten seconds earns reuse; one buried under a wall of text stops getting used, however good the material inside it is.
Connect your AI to real tools
A plug (an MCP server) reaches into a real account and pulls live numbers or takes real action, ahead of working from a description. Each one authenticates through the tool's own login screen, never a token copied from a stranger. Four real install commands:
Start with one, add the rest once it earns its place. More plugs worth having, matched to what they actually connect:
| Plug | What it reaches |
|---|---|
| Klaviyo | Email flows, campaigns, segments, live open and click data |
| Google Ads | Search campaigns, keywords, real spend and conversion numbers |
| Airtable / Notion | Your own trackers, content calendars, client databases |
| Supabase / Stripe | Your own product's real usage and revenue data |
A plug that connects but returns nothing useful is worse than no plug, it looks trustworthy and isn't. Test a new one before relying on it:
Load in proven ad creative formulas
Roughly 70 to 80 percent of ad performance comes from the creative itself, ahead of any audience setting. Four formulas cover almost every angle worth testing:
Generate real angles against a real offer:
Test one variable at a time or the result teaches you nothing. Five hooks against the same offer, same audience, same budget, same seven days, tells you which hook wins. Five hooks plus five audiences plus three offers, all launched together, tells you nothing, a win or a loss could belong to any one of them.
Under 20 conversions on a variant is noise. Hold out for 50 or more conversions per variant on image and text tests, 30 or more on video, engagement metrics arrive faster there, before calling a winner and killing everything else.
More hooks worth stealing, once PAS, testimonial, and question have run their course:
Four finished templates built to these formulas sit inside the Meta ad creative pack below.
Want both of these AI agents in your own folder?
Prompt Engineer and the Meta Ads Creative Engine, sent to your inbox in one email.
Get your two free AI agentsGenerate assets at scale
Take the strongest angle from step five and produce the real image or video. A strong Higgsfield prompt names the subject, the setting and light, the feel, space for a headline, and the shape:
Generate several angles in one pass, so the first test starts with real options instead of a single guess. Match the shape to where the ad runs before generating anything, regenerating for the wrong shape wastes a full round trip:
| Placement | Shape |
|---|---|
| Feed, image or carousel | 4:5 vertical |
| Stories and Reels | 9:16 full vertical |
| Square placements, some carousels | 1:1 |
For video, the ad lives or dies in the first three seconds, so generate the hook frame deliberately instead of hoping the first second looks right:
Start on the fast, cheap generation model to prove the composition and motion work before spending on anything premium. Reserve a higher-cost model for the concept that has already won a test, ahead of the concept you're still guessing at.
Set guardrails before you launch
Before an AI changes anything in a real account, publishes anything public, or spends a dollar, it shows the plan and waits for a yes. Set the rule once:
New campaigns launch paused. A human turns them on. The standing rule above covers publishing. Layer three more, specific to where real mistakes actually happen:
Cap any single automatic increase at 20 percent, and require 48 hours between increases. An AI that can raise budget without a ceiling can also raise it by 500 percent on a bad read of the data.
A drafted email or message always waits in a drafts folder for a human send, never an auto-send, even for something as small as a reply.
One account per session. An AI connected to five ad accounts at once is one wrong assumption away from changing the wrong client's budget.
Make the AI check its own work
A confident answer and a correct one are different things. Ask for evidence instead of a summary:
A verification claim is only as good as the independent check behind it. Match the claim to a real check, ahead of a restatement of what the AI just said it did:
| Claim | Real check |
|---|---|
| "The campaign is live" | Read the status back from the ad account itself, paused or active |
| "The email sent" | Check the email platform's own send log for that exact message |
| "The page is up" | Fetch the URL and confirm a real 200 status code |
| "The build passes" | Run the build command and read the actual exit code |
An AI that skips this step reports something done because the words it generated described doing it. A description of an action and the action itself are different things, and only one of them shows up in a client's account.
Scale by the numbers
The numbers that actually matter for a typical online offer:
| Metric | Good | Great |
|---|---|---|
| ROAS (return on ad spend) | 2x | 3x+ |
| CPA (cost per sale) | under $50 | under $30 |
| CTR (click rate) | 1% | 1.5%+ |
| Hook rate (past 3s of video) | 25% | 35%+ |
Kill a creative once it holds under a 0.5% click rate past a thousand views, or spends twice the target cost per sale with nothing to show. Scale a creative once it beats the target for three straight days on 50 or more results, raising budget 10 to 20 percent every two to three days. A big jump resets what the platform learned and can wreck a real winner.
Different objective, different bar. A lead-gen business should never be graded against ecommerce ROAS targets:
| Metric, lead gen | Good | Great |
|---|---|---|
| Cost per lead, content offer | under $40 | under $25 |
| Cost per lead, demo request | under $150 | under $100 |
| Lead to qualified rate | 10% | 20%+ |
Watch frequency, not just the headline metric. A creative sitting above a frequency of 3 with softening click rate is fatiguing, well ahead of any drop in the daily number telling you first. Split spend roughly 60 to 70 percent on prospecting a broad, cold audience, 20 to 30 percent on retargeting the warm one, and hold back 10 to 20 percent purely for testing new angles, so this week's testing budget never eats next month's scale budget.
Your tool stack, in one line each
Every plug uses your own login through the tool's own screen. Nothing gets pasted in from a stranger's video.
Where most AI marketing attempts fall apart
Five folders, and the habit that runs them
Inside your notes app, this is the whole starting structure. Folders are cheap; add more later.
The habit: if a fact matters, it goes in the vault. After a call, paste notes under People. After a campaign, write down what happened under Results. Ask the vault a question later and the AI answers from what actually happened in the business, ahead of a guess.