Jonathon's AI Wiki

The Marketing Engineer — A Growth-OS Repo Run by Agents (Greg Isenberg)

9 min read

Source: raw/Marketing_Engineer_-_The_1M_Job_with_AI_Agents.md — Greg Isenberg, Startup Ideas podcast, solo episode, youtube.com/watch?v=8ZC1G1ezN5o, YouTube. Auto-caption transcript fetched 2026-09-29.

Transcription normalization

The captions render Grok Bot as “Grockbot,” Codex as “Codeex,” fal as “foul AI,” Ahrefs as “Hrefs,” and banned language as “band language.” All are corrected below. Grok Bot is xAI’s agent-team product; see Grok Bot.

Greg argues that the next high-value marketing hire is a “marketing engineer”: “the person who can do a whole marketing team work with AI agents,” whose job is to turn scattered market signal into pipeline. The practical core is a “growth repo” (a structured folder that holds the company’s marketing memory), a seven-field job spec for every agent, six systems built on top of the repo, and a 30-day plan for learning the role on one real company. The salary claims are his opinion. The repo layout and the job spec can be used as they are.

Key Takeaways

  • Definition. “The person who turns market signal into pipeline using AI agents, data, code, taste.” Other names in use: “forward deployed marketer” and “AI growth operator.”
  • Greg’s history of marketing eras:
    • traditional: making people care;
    • digital: acquiring customers through channels you can measure;
    • growth hacking: product and data loops;
    • marketing engineering: “a marketing system that keeps learning.”
  • The salary claim is opinion. He expects “a 250k, 500k, a million dollar job” and calls a million “conservative.” He gives no data.
  • The problem it solves is scattered learning.
    • Sales, support, product and marketing each hear a different version of the market, and “everyone walks into the growth meeting with a slightly different version of reality.”
    • Most AI work happens in “random chats,” where the output “just disappears” and next week “the AI is starting from scratch again.”
  • Prompts get better once the repo exists. Instead of “write me 10 LinkedIn posts,” the prompt is: read the customer-truth file, the founder-voice file and “the last five posts that drove qualified replies,” then draft five posts about the pains buyers mentioned this week.
  • Every agent gets a job spec written “like I was hiring a person”:
    1. data source;
    2. when it runs;
    3. what it filters out;
    4. expected output (“here’s what good looks like”);
    5. what needs human approval;
    6. the metric that matters;
    7. where results are written, “so the system gets smarter next time.”
  • Worked example: a competitor-engager agent.
    • Every weekday morning it checks 20 LinkedIn accounts and pulls the people who commented on new posts.
    • It enriches them, drops bad fits, and drafts 10 messages, each tied to the post that person engaged with.
    • It writes the drafts to a file for approval.
    • The metric is “positive replies from qualified accounts”; “messages sent is activity.”
    • This is the same agent built step by step in Cody Schneider’s LinkedIn-signal agent.
  • Train agents like new hires, and turn each correction into a rule. Start small, watch, correct, “add the correction to memory,” then widen the scope.
    • If a first line sounds fake, add a rule to the repo.
    • If intros sound generic, give “three good examples and three bad ones.”
    • If a claim has no evidence: “every insight needs a quote or a link or a source.”
  • An SEO agent, beginner versus marketing-engineer version. The beginner asks for “a blog post about a keyword.” The marketing engineer’s agent does this:
    1. checks Google Search Console;
    2. pulls keyword data from Ahrefs or Semrush;
    3. checks the CMS for existing coverage;
    4. ranks topics by volume and buyer intent;
    5. researches what already ranks;
    6. adds the founder’s point of view;
    7. drafts the post with a meta title and internal links;
    8. sends it for approval.
  • Greg’s tool split (he calls it “just the way I’m thinking about it”):
    • Grok Bot as “the growth operating system that lives close to the internet,” because it is connected to X. One bot watches competitors, one watches customer language on X and Reddit, one watches niche creators for formats to test, and one watches ads and landing pages.
    • Claude and Codex build the repo, landing pages, scripts and small internal tools.
    • Hermes-style workflows for scheduled operations “with memory and approval.” Examples: build a market brief every Monday; review experiments every Friday; when a new batch of sales calls lands, pull the objections and update the positioning file. See Hermes Agent.
    • Creative models (fal, Higgsfield) for ads, thumbnails, mockups and video concepts.
    • Local models when the data is sensitive (private transcripts, regulated notes, pricing plans) or cloud processing is too expensive.
    • “The tools are going to keep changing, but the workflow is the thing to actually learn.”
  • Where the value is. “The agents are going to be a commodity at some point. Your judgment about what to point them to is the moat.”

The Growth Repo

Greg describes it as “a GitHub repo… or honestly just a structured folder,” named something like growth-os:

FolderContents
Customer truthSales call notes, support tickets, churn notes, interviews, live product feedback
Content engineFounder voice guide, winning hooks, scripts, notes on what performed
Outbound engineICP, account research, trigger events, approved angles, banned language (“AI outbound gets weird fast”)
Creative testingAd angles, landing-page tests, hooks, offers, results
AgentsJob definitions for each AI worker

The Six Systems, With His HVAC Example

His example company sells software to commercial HVAC contractors. The sharp angle he is after is “stop losing replacement revenue after every service call,” not “run your HVAC business better.” All of the numbers in this section are illustrative.

  1. Customer truth.
    • Output: a file called what the market is telling us.md, updated daily or weekly.
    • Inputs: sales calls, support tickets, churn notes, CRM notes, Stripe movement and social data.
    • It shows what changed, with “quote snippets, ticket links, event counts.”
    • A weak memo says “customers want better collaboration.” A strong one says: “five sales calls this week mentioned emergency dispatch, but the calls that actually converted all talked about missed follow-up quotes after the tech left.”
  2. Founder content engine.
    • Record the founder talking to customers, mine podcasts for the strongest ideas, and track which hooks hold attention.
    • One insight becomes several assets: a founder post, a short video, a landing-page line (“every completed job should create the next quote”), a cold-email angle and a lost-revenue calculator.
  3. Outbound signal engine.
    • Start from timing, not a spreadsheet of names: who raised money, who is hiring for the problem, who posted about the pain.
    • HVAC signals: hiring dispatchers, opening new locations, getting bad reviews.
    • The agent researches accounts, drafts angles and sends them to a human for approval.
  4. Creative testing engine. One offer becomes 20 hooks and 10 ad angles, and the results are recorded, so creative becomes a learning system instead of a treadmill.
  5. AI search visibility.
    • The question: can ChatGPT and other assistants understand the company?
    • Agents pull that data, then create content and optimize the site so it gets cited.
    • Greg cites Sam Altman as saying ChatGPT has a billion users; this is secondhand.
  6. Growth cockpit.
    • A weekly view of what content worked, which campaign started real conversations, which objection came up again, the test win rate, what competitors did, which pain is getting louder, and what to test next.
    • Example memo: the lost-replacement-revenue angle “drove fewer clicks than the dispatch angle, but twice as many demo requests from owners with more than 20 techs.”

Ways to Make Money (Greg)

  • In-house. The role sits “directly next to revenue,” which is why he thinks it will pay so well.
  • Consulting. Embed with a company for 30, 60 or 90 days, build one system, and charge “5, 10, $30,000 a month.”
  • Productized service. Repeat one narrow offer, for example: outbound signal engines for vertical SaaS; founder content engines for B2B CEOs; customer-truth repos for seed-stage startups.
  • Software. Build the same system by hand for 5 to 10 companies, then turn whatever repeats into a product. He would start with services.

Try It

  • Build the “almost painfully simple” first version.
    • Make a growth-os/ folder with files for customer truth, founder voice, experiments and agent jobs.
    • Paste in 20 real customer notes or call summaries.
    • Give the agent one job: “tell me what’s changed, show me the receipts, suggest one marketing test that could create pipeline this week.”
    • Build one thing from the answer.
  • For one agency client, such as a dental practice, the customer-truth inputs would be new-patient call notes, reviews and objections heard at the front desk.
  • Write the seven-field job spec before automating anything. Make “where results are written” a file in the repo, so corrections build up in one place.
  • Run Greg’s 30-day plan:
    • Week 1: audit one real company. Study the website, offer, ICP, founder content, and sales calls and tickets if you can get them. Produce a market map: who the customer is, what pain they describe, the words they use, what they buy instead, where the funnel leaks, and what you would test first.
    • Week 2: create the repo and the first what the market is telling us.md.
    • Week 3: build one system. “One working system is going to beat five half-built ones.”
    • Week 4: measure results — replies, meetings booked, conversion, whether the founder sounds sharper. His example case study, “75 targeted messages got nine warm replies booked three calls,” is hypothetical, not a reported result.

Open Questions

  • Pay figures. The salary and consulting rates are Greg’s estimates, with no data behind them.
  • The HVAC numbers are all illustrative; nothing in the episode is a measured result.
  • The repo is described, not shown. The transcript gives no repo, prompt files or agent specs, only folder names as spoken. He says “five of those files” for the minimum version but names four.
  • How it compares to agency work. The episode doesn’t compare this model with an agency running the same systems for several clients.

Graph View

Backlinks