Finding signal on Twitter is more difficult than it used to be. We curate the best tweets on topics like AI, startups, and product development every weekday so you can focus on what matters.
Marketing engineers are the new 10x employee, and most companies don't have one yet.
If you're growing a startup, a business or a side hustle, becoming one or hiring one is the biggest lever you have right now. It's the most exciting time to be in marketing since Google launched AdWords. Maybe ever.
For 20 years, marketing ideas were easy and shipping them was hard. Sales knew why deals closed. Support knew why customers left. But every idea had to survive a brief, a designer, a developer, an analyst and two meetings, and most died in a Slack thread.
That bottleneck just broke!
Today one agent reads last week's sales calls in ten minutes. Another turns what customers said into a landing page. Another launches thirty ads and kills the losers while you sleep. Another tells you on Monday which campaign brought in customers who paid. And the models are now good enough that if your idea is good, the output will be too.
A marketing engineer wires those agents into a machine. Part marketer, part builder. And the best ones often come from outside marketing: founders, engineers, operators, solo builders So grab your coffee, grab your tea because here's the full playbook (15-20 min read):
How we got here In the Mad Men era, from the 1950s through the 1990s, marketing meant print, TV, radio and billboards. The job was big ideas and big budgets. A campaign cost millions, took months to produce, and you found out whether it worked a quarter later if you were lucky. A great marketer got a handful of real shots a year, so the whole profession ran on taste, persuasion and gut feel.
Then Google launched AdWords in 2000 and Facebook launched ads in 2007, and suddenly every click could be counted. The digital marketer was born. The job turned into managing channels like SEO, paid search, email and display. Marketers lived inside dashboards, and the best ones were whoever understood the platforms most deeply. You could run a few experiments a month.
In the 2010s came the growth hacker, and I'll never forget it. Sean Ellis coined the term in 2010, and the decade's best tricks became folklore. Hotmail put “P.S. I love you. Get your free email at Hotmail” at the bottom of every message. Dropbox gave you and a friend free storage for every invite. Airbnb let hosts crosspost their listings to Craigslist. The growth hacker treated marketing like product, with funnels, loops and A/B tests. The catch was that every good idea needed an engineer, and engineers belonged to the product roadmap. Growth hackers spent half their lives writing tickets and waiting in line. A few experiments a week was a great week.
When ChatGPT showed up at the end of 2022, we got the AI assisted marketer. Copy came faster and images came faster, and everything else stayed the same. Same org chart, same meetings, same six week cycle, just with better first drafts. This was kinda like the AI copilot era. Most marketing teams are still sitting right here.
The marketing engineer is what happens when the person with the idea and the person who can build it become the same person. One marketer researches the customer, builds the page, generates the creative, wires up the data and launches the test. The unit of work shifts from a campaign to a system that keeps producing campaigns, and dozens of experiments a day becomes realistic for one person.
Zoom out and the whole history of marketing is one trend. The cost of an experiment keeps collapsing. Whoever runs the most good experiments learns the fastest, and whoever learns the fastest wins.
Why this matters so much
Every company is sitting on a pile of gold it rarely uses (I know it's hard to believe but it's true).
Sales hears the same five objections every week. Support knows exactly which promises lead to refunds. The founder has strong opinions about the category that live in their head and a few podcast clips. Analytics knows which pages bring in customers who stick around. All of it lives in different tools, owned by different people, and it reaches a campaign maybe once a quarter.
Meanwhile the small ideas that would put that knowledge to work keep losing to the big launch. A pricing calculator for one niche. A campaign built from yesterday's sales calls. A note to forty old prospects on the day the feature they asked for finally ships. Each one needed a designer, a developer, an analyst and two meetings, so each one sat in the backlog forever.
A marketing engineer can do every one of those in an afternoon. Specific marketing got cheap, and when specific gets cheap, the company that understands its customer best starts winning. Agencies that sell hours will feel this first. Big marketing teams will start to look like committees. Founders and solo marketers will suddenly have the output of a whole department.
Note: My partner just launched Boring Funnels, the idea being that the biggest opportunity is helping businesses turn what they know about their customers into clear offers, useful landing pages and follow-ups that lead to sales. A marketing engineer can build the whole thing, see where people drop off and keep improving it as more customers come through.
Here's exactly how to build it.
Step 1: Set up your desk
Everything starts with an agent workspace. Claude Code, Codex and Cursor are the common ones. The easiest way to think about these tools is as a very fast junior employee who lives inside a folder on your computer. It can read every file in that folder, write new ones, run code, and use any tool you connect to it.
That last part is where the magic is. Tools connect through something called MCP, which is basically a universal plug for AI. HubSpot, Stripe, Notion, Slack and hundreds of other tools have MCP servers, so your agent can read your CRM or pull last week's revenue the same way you would. When a tool only has an API, the agent can usually write the code to call it.
You'll also want your agents running on a schedule while you sleep. A year ago that meant stitching things together in a workflow tool like Zapier or n8n. Today you can just ask your coding agent to build it. Say “run this every Monday at 7am, post the report to Slack, and wait for my thumbs up before sending anything,” and it writes the scheduled job, the Slack message with an approve button, and the code that only acts once you tap it. Or use an always-on agent like Hermes or OpenClaw that already runs in the background. Either way, it's your code, in your folder, and you can change any part of it with a sentence.
Step 2: Write the brain
Here's the mistake I see over and over. Someone gets excited, spins up ten agents, gets back a mountain of generic slop, and decides AI marketing is overhyped.
The agents were fine but the context was thin.
It's like hiring a brilliant marketer and skipping onboarding. They'll write something smart about your industry, and it'll sound exactly like your competitors. So before you build a single agent, build the brain: a folder that explains your business the way you'd explain it to a sharp new hire on their first day.
Here's the structure I'd use:
Six files do most of the heavy lifting.
1/ START-HERE.md is the one page every agent reads first. It says what the company does, who it serves, what this quarter's priorities are, and where everything else lives.
2/ company/customer-fit.md describes who buys, who stays, and who you can serve profitably. Be painfully specific. “Shopify stores doing $1M to $10M a year, where the founder still does the books and has been burned by a bookkeeper before” gives an agent something to work with. “SMBs” gives it very little.
3/ company/offers.md lists what you sell, what it costs, what you're allowed to promise, and what needs approval before anyone says it out loud.
4/ brand/voice.md is where most people get lazy. “Friendly but professional” is useless to an agent. Paste in three emails or ads you loved, with a sentence on why each one works, and two you rejected, with the reason. “We killed this one because it promises results in 30 days and we can't guarantee that.” Agents learn far more from examples than from adjectives.
5/ brand/founder-opinions.md is your unfair advantage. Every competitor has access to the same models. Only you have your founder's takes on the category, the customer and the competition. Spend two hours interviewing your founder, or yourself, and write down every strong opinion. “Most agencies in our space overcharge for onboarding.”“Our best customers come from accountant referrals.” Those opinions are what make the output sound like you.
6/ decisions/failed-campaigns.md holds the lessons that cost real money. “The free audit attracted companies too small to afford us.”“LinkedIn brought lots of leads and very few closes.” One line like that saves an agent from repeating a $20,000 mistake.
One rule keeps the brain healthy: agents can add guesses to evidence/, but only a human moves something into company/. Otherwise one agent's hunch quietly becomes company policy.
Step 3: Connect your data
Next, give the brain eyes. Connect the places where customer truth lives: call recordings in Gong or Fathom, support tickets in Intercom or Zendesk, deals in HubSpot or Salesforce, behavior in PostHog or Google Analytics, and revenue in Stripe.
Start with read-only access everywhere. An agent that can read your CRM is useful on day one. An agent that can edit your CRM is something it earns over time.
Then do the boring thing almost every company skips. Give every campaign an ID and carry it everywhere. Put it in the UTM tags on every link, in the name of every ad, in the name of every email sequence, and in a field on every deal in the CRM. Then keep one simple table, in Supabase or even a spreadsheet to start, with a row per campaign: ID, spend, leads, qualified calls, deals, revenue and churn.
It sounds unglamorous, and it's the most valuable thing in the whole setup. Most companies can tell you clicks by campaign and revenue by customer, and very few can connect the two. Once you join those tables, your agents can answer the question that actually matters: which marketing brought in customers who paid and stayed?
Step 4: Build your first agent
Build one agent and get it great before you build a second. Six agents talking to each other on day one is how you end up spending your weekend debugging a group chat.
Your first one should be the customer language agent, because everything else feeds off what it finds. Its job is to read what customers actually say and pull out the exact words they use to describe their problems.
Export your last 20 to 50 sales call transcripts into evidence/calls/ and give it an instruction like this:
Read every transcript in evidence/calls/.
Then read the report next to your own memory of those calls. The first version will be about 70% right. It'll probably over-weight one chatty customer, or name the groups in marketing speak. Every time you spot a mistake, fix the instruction and save the fix in decisions/ so it sticks. After three or four rounds it'll read calls better than most people on your team.
Here's the kind of thing it finds. Say you sell scheduling software to plumbers and HVAC companies. The agent notices owners keep saying some version of “I'm tired of taking calls during dinner.” That one line is your ad, your headline and your sales opener. It beats “automated scheduling” every time, because people buy the situation more than the feature.
Once it's reliable, put it on a schedule. Every Monday it reads last week's calls and tickets and drops a fresh report into evidence/customer-language/. Now your entire marketing operation runs on what customers said last week, instead of what someone guessed last year.
Step 5: Add the other five agents
Once your first agent is solid, add the rest one at a time. Each one gets its own folder in workflows/, its own instructions and its own schedule.
The buying trigger agent watches for moments that create demand overnight. A company opens a second location. A brand starts selling wholesale. A startup raises a round and posts five sales jobs. A tool you integrate with ships something new. The easiest way to build this is in Clay. Pull a list of accounts that match customer-fit.md, then add columns that check for each trigger, like new job postings, funding news, new locations or new pages on their website. Apollo is handy for the contact data.
For every account with a fresh trigger, the agent checks the CRM to skip anyone your team is already talking to, then finishes one sentence: “This event creates this problem, and our product solves it in this specific way.” If the sentence comes out mushy, the account waits. If it comes out sharp, the agent drafts a short email that leads with the event and saves it as a draft for you to review. That one-sentence filter is what makes outreach feel like good timing instead of spam.
The search agent finds the questions people ask right before they buy and builds the best page on the internet for each one. Feed it three sources: your Search Console data, the questions that come up on sales calls, and what ChatGPT, Claude and Perplexity say when you ask them about your category. That last one is new and wildly underrated. Every week, ask the AI engines the questions your buyers ask and record which companies and sources they recommend. It shows you what the models trust and where you're invisible.
Then build pages that are more useful than anything else out there, because AI search rewards the most useful page more than the most optimized one. A commercial kitchen supplier that builds a guide to opening your first restaurant, with a calculator for what equipment fits your space and budget, gets ranked, gets cited by the AI engines, and gets bookmarked by buyers who come back when they're ready. Your agent can write that calculator in an afternoon and publish it on Next.js and Vercel, or on Framer if you want speed.
The creative agent makes and tests ads. Give it the latest customer language report, the approved claims from offers.md, real product footage and screenshots, and a folder of ads you love. Ask it for different reasons to buy, instead of fifty versions of the same headline: one ad that walks through a worked example, one that tells a customer story, one that's a raw demo of the product. Image models like Nano Banana and video models like Veo make the production almost free.
Name every ad with what it's testing, like BK014_dinner-angle_ugc_v2, so the campaign ID, the idea and the format live right in the name. A month later, “the dinner angle beat the price angle” is a lesson you can use everywhere. “video_final_7 won” is trivia.
The agent uploads ads through the Meta Ads API in a paused state, so a human hits go. Then a small script checks results every morning and pauses anything under 1% CTR after a set amount of spend. Decide your evidence bar before you launch: how much spend, how many days, and what counts as a win. Skip that step and 200 ads just gives you 200 inconclusive experiments.
The closed-lost agent is the one that pays for everything else. Pull every lost deal from your CRM along with its notes and call transcripts. “Not interested” is a lazy summary of a real conversation, so have the agent reread each one and record the actual reason plus a reopen condition. “Went with a competitor, contract renews in March.”“Needed SSO.”“Champion left the company.”
Then it watches for those conditions. The day your changelog says SSO shipped, every prospect who asked for it gets a personal draft that quotes their original call. When the renewal date gets close, you show up. When the champion lands at a new company, you congratulate them and reintroduce yourself. Respect opt-outs and contact preferences, and you'll find most companies have a pile of unfinished revenue sitting in their own CRM.
The performance agent reads your campaign table every Monday and writes you a short memo, the way a sharp analyst would. “The profitability review page converted at 4%, and half those leads were too small to serve. The dinner ad has the lowest cost per qualified call. Retargeting has spent $3,000 and produced two sales conversations.” Every line ends in a decision for you: keep, kill, change or test.
Wire it all the way through to retention, refunds and profit. A campaign that brings in cheap customers who churn in sixty days looks amazing on a dashboard and quietly loses you money. This is the agent that catches it.
Step 6: Put the guardrails in the tools
The fastest way to get burned is to trust a sentence in a prompt. “Always ask me before sending” is a polite suggestion to a model. Real guardrails live in the tools.
Give every agent the smallest set of permissions it needs. The outreach agent gets draft-only access to email. The ads agent creates ads in a paused state, and budgets stay with you. Approval steps live inside the code itself, like a Slack message with an approve button that the action waits on. Use plain code for exact rules like spend caps, suppression lists and deduping, and save the model for judgment calls like whether an account fits or whether a message sounds like you.
Then write each agent a short contract, like a job description with guardrails. Here's one for the buying trigger agent:
“Done” means it checked every source within its limits and told you what broke. An agent that says “this source failed today” is worth ten that quietly skip it.
Step 7: Run a campaign end to end
Here's what it looks like when all of it works together. Picture a bookkeeping firm that wants more ecommerce clients.
On Monday, the customer language report shows that store owners keep saying the same thing: “I can't tell if I actually made money after returns, shipping and fees.” Eleven different customers said it last month. You walk it over to the person who delivers the service and ask whether the firm can really solve this. They say yes, and that it's the first thing they fix for every new client. You approve an offer in offers.md: a free profitability review for stores doing $1M to $10M a year.
The campaign gets an ID, BK014, and everything from here carries it.
On Tuesday, the creative agent makes three angles from the customer's own words: where your profit disappears, a sample monthly report, and a short story from a real client. The search agent builds a page with a simple profit calculator that uses numbers the firm has checked. The closed-lost agent finds 60 old prospects who said “maybe after year-end” and drafts each of them a personal note. The buying trigger agent flags stores that just launched on Amazon, because multichannel selling makes the books messier.
On Wednesday, you review everything in one sitting. You cut one ad that promises savings the firm can't guarantee, approve the rest, and launch a $2,000 test.
Two weeks later the performance agent connects the spend to booked reviews and the reviews to signed clients. The calculator page converts best. The Amazon trigger list books the most calls. It also reminds you the sample is still small.
You'll notice that he customer's original words survived every step. Normally a sharp insight goes through a research summary, a brief and five rounds of edits, and comes out sounding like every other bookkeeping ad on the internet. Here the quotes traveled with the campaign folder, so every agent saw them.
Step 8: Run the weekly loop
Once the machine is built, the job looks a lot more like editing than making.
Monday you read the performance memo and make the keep, kill and change calls. Tuesday and Wednesday you build, fix the workflows that produced weak work last week, and launch new tests. Thursday you get on calls with sales and support, because that's where next week's best ideas are hiding. Friday you review what the agents made and update the files.
The heart of the whole thing is a review queue. Every item in it should take under a minute to decide. “Approve this revised offer. It changes these two ads and this email.” Show the change, what it touches, and the evidence behind it.
And the system only gets smarter when you change its inputs. Every time you reject something, write down why and put it where the next run will see it. Rejected an ad because it pulls in businesses too small for you? Update customer-fit.md. A prospect misread an offer? Add a check for it. Keep an experiment log with what you tested, why, what happened and what you'll do next. Keep ten example inputs for each agent with the output you expect, and rerun them every time you change an instruction, so you know right away if something got worse.
If you find yourself spending all day fixing agent output, the problem is upstream. It's almost always the brain or the contract.
The creative engine
Here's what most people still haven't caught up to: creative is the new targeting.
Meta rebuilt how it decides which ads to show. Its new system, called Andromeda, finished rolling out globally in late 2025, and it reads the ad itself to figure out who should see it. Your image, your hook and your script now do the job that interest targeting used to do. One side effect is brutal for lazy advertisers. Andromeda groups ads that look alike and treats them as one, so fifty versions of the same headline get you roughly the reach of a single ad. And winning ads now burn out in two or three weeks instead of six.
So the job is a steady stream of truly different ideas. Until recently that needed a production team. Now it needs a creative engine, and here's how I'd set one up.
Start with the words. Load voice.md and the latest customer language report into a frontier model like Claude or GPT and have it write scripts for 8 to 12 distinct concepts: a pain point, a worked example, a customer story, a founder rant, a comparison, a myth you're busting. Give each concept two or three hooks. The strategy lives in this step, so spend your time here.
One shortcut that helps a lot: start from a funnel that's already working.
My partner hasn't publicly talked about it much but @BoringMarketer just launched Boring Funnels, which breaks down a real, live funnel every week, from the ad to the last email, and turns each step into a prompt you can hand to your coding agent. It's a much better starting point than a blank page.
Then make the images. Nano Banana, GPT Image and Midjourney now produce product shots, lifestyle scenes and static ads that hold up on a phone screen. The trick is to feed them your real product photos so the product stays accurate, and let the model change everything around it.boringfunnels.com Then make the video. Veo, Sora, Kling and Runway turn a script or a still image into a short clip, and Higgsfield adds camera moves that look like a real shoot. For talking-head ads, Arcads, HeyGen and Creatify generate AI actors reading your script in dozens of languages. ElevenLabs handles voiceover, and CapCut or Captions handles the edit.
Use AI actors to test, then pay humans to win. This is the inside sauce on AI UGC. AI actors let you test twenty scripts in a day for the price of one creator video. Once you know which script works, a real person saying it usually performs even better, builds more trust, and keeps you on the right side of platforms that require AI content to be labeled.
Let the agent run the loop. It launches each concept paused, a human hits go, a script pauses the losers after a set amount of spend, and every Monday the performance agent tells you which concepts won and which hooks to try next. Your job becomes picking ideas and judging taste, and the machine handles the volume.
The marketing engineers winning on paid social right now are shipping dozens of truly different concepts a month, and they're doing it without a production team.
The social engine
Paid gets all the attention, but organic social is where marketing engineers quietly build an unfair advantage. It's free distribution, and it's the cheapest creative testing lab ever invented.
Treat organic as your test kitchen. Post ten hooks as organic TikToks, Reels or X posts, see which ones the algorithm picks up, and only then put money behind the winners. You find out what works for free, and your ad budget only goes to ideas that already proved themselves.
Clip everything. One long piece of content, like a podcast, a webinar or a founder rant on a walk, can become twenty short clips. Tools like Opus Clip find the best moments and caption them, and an agent can schedule them across every platform. This is how most of the biggest podcasts grow.
Use formats that scale without a face. Slideshows and carousels on TikTok and Instagram, built from your customer language report, are some of the easiest posts to make and some of the most saved. An agent writes the slides, a model makes the images, and you approve the batch on Sunday.
Listen before you talk. Point an agent at Reddit, X, TikTok comments and your competitors' reviews, and have it report what people are complaining about, asking for and getting excited about each week. Then have it draft helpful replies for threads where your product truly helps, and post them yourself. On Reddit especially, a useful answer from a real person beats any ad, and a spammy one gets you banned.
Put the founder to work. When every brand can generate content, the scarcest thing on the feed is a real person with real opinions. Turn founder-opinions.md into a content engine: the agent drafts a post a day from your takes, and you edit it into your voice and hit publish. AI makes the founder more valuable, because they're the one thing nobody else can generate.
Two rules keep this clean. Label AI-generated people and scenes wherever the platform asks, and keep a human on the publish button.
How to become a top 1% marketing engineer
Once the basics run on their own, this is where it gets fun.
The first is building decision tools. A packaging company builds a calculator that shows which box size cuts your shipping bill. A software company builds a migration checker that scans your setup and shows exactly what moves over cleanly. These used to be six-month engineering projects, so they rarely happened. Now a marketing engineer can ship a scrappy version in a week, and buyers bookmark it and come back right when they're ready to buy.
The second is marketing to the people who influence the purchase. The accountant who recommends software to forty clients. The agency that picks tools for its customers. The IT person who gets asked “is this safe?” Build a campaign just for them, with exactly the material they need to say yes.
The third is fixing acquisition inside onboarding. Sometimes the ads are working fine and new customers churn because the first week is confusing. An agent that reads onboarding tickets and finds where people get stuck can do more for growth than your next ad set.
The fourth is giving every channel its own job. Search answers questions. Social starts conversations. Email shows up at a specific moment. When every channel repeats the same message, you're paying for the same impression five times.
The stack
Here's the full toolkit, by job. Tools change every few months, so learn the categories and swap the names as you go.
Agent workspace: Claude Code, Codex, Cursor, Google AI Studio
Search and AI visibility: Ahrefs, Semrush, Google Search Console
Ads: Meta Ads API, Google Ads API
Email and outbound: Customer.io, Resend, Instantly
Web: Next.js on Vercel, Framer, Webflow
Measurement: PostHog, Stripe, Supabase
What you need to learn
The job is half marketer and half builder, and you need both halves.
The marketer half is the stuff great marketers have always known. How to run a customer interview and hear the phrase that matters. How to position a product so the right people feel it was made for them. How to build an offer that's a real reason to buy, as opposed to a discount in a costume. How to write a headline that stops a scroll. How each channel decides who sees what. And enough statistics to know when a result is real and when it's a fluke.
The builder half is newer, and it's much easier to learn than people think. You need to write context that an agent can actually use, which is the most underrated skill on this whole list. You need to read code well enough to judge whether it does what you asked, because the agent writes it and you own it. You need to understand how APIs and MCP connect tools, write basic SQL to pull your own numbers, ship a page, use git, write test cases, and know roughly what each agent run costs you.
The fastest way to learn the builder half is to ship one ugly landing page this weekend with an AI coding tool. One deployed page teaches you more than ten courses.
Where this goes
If I had to start this week, here's what I'd do. If you're a marketer, build the customer language agent and run it on your last 20 sales calls. If you're an engineer, sit in on five sales calls, then build a calculator for one customer segment. If you're a founder, block two hours this weekend and write founder-opinions.md, because every agent you build after that gets sharper.
Keep publishing, outreach and spending behind approval until each workflow earns your trust. Get one workflow reliable before you add the next. A confused workflow running every morning just gives you a reliable supply of confusion.
For most of the history of marketing, the winner was whoever had the biggest budget and the most people. I think the next decade belongs to whoever understands their customer best and can act on it fastest. My guess is that by 2028 the best marketing teams will be one or two marketing engineers running dozens of agents, and they'll outship departments ten times their size.
Most of those people are sitting in growth, content and ops roles right now. They're one weekend of building away from becoming one.
If you want to go deeper, I did a full episode on marketing engineers on the podcast: https://youtube.com/watch?v=8ZC1G1ezN5o DM me if I can ever be helpful. Can't respond to everyone but will respond to some!
I'm rooting for you.
Greg Isenberg
Btw, ways we could work together:
Hire LCA. We're the leading product design firm for AI. If you want your AI product to look and feel world class, reach out. Strategy, design, product, engineering, the whole thing.
Use Ideabrowser.com. Find your next startup idea, backed by real data. Over 200k founders use it.
Try Boring Funnels. My partner just launched Boring Funnels. Every week it breaks down a real funnel that's working and turns it into prompts for your coding agent.
