Marketing Engineer: Handing Marketing Workflows Over to AI Agents
Shann³ shared his marketing engineer framework: in an agent development environment, he uses a knowledge layer and a data layer to support workflows and achieve marketing automation. He provides the specific directory structure, API connections, workflow design, and looping methods.
A new job title is becoming popular in marketing circles: marketing engineer. It sounds like equipping a marketer with a programmer, but in reality it's one person doing two jobs.
Shann³ recently published a long post that breaks down his framework in detail. He puts 70% of marketing work into an Agent Development Environment (ADE), using multiple Agents to collaborate on research, planning, content production, and landing page building. This image shows the overall architecture:
What is a marketing engineer? Simply put, it's building the system behind marketing. Use Agents, connect data and business context, and automate marketing work. Marketing knowledge determines how the system works: you decide what the Agent needs to know, which sources to use, how to check outputs, and which steps need your involvement.
Before starting, understand three terms:
- Model: the AI chosen for a task, such as research, planning, or writing.
- Harness: the software around the model that gives it files, tools, and code execution permissions.
- ADE (Agent Development Environment): the interface where you work directly with Agents, review their actions, and guide their work.
Shann³'s main workspace is Orca, but recently he has been using Claude more, and he is also using Grok bot and Hermes. He mentions a key point: don't over-engineer. He once built four Agents for a content system, then found that improving context was more useful than adding Agents.
How to set up the workspace
Each project needs a place to store context, run workflows, and save outputs. He organizes by marketing vertical, with each vertical having its own directory, and shared knowledge placed separately.
markdown
marketing/
├── shared-knowledge/
├── content/
│ ├── knowledge/
│ ├── work/
│ ├── workflows/
│ ├── skills/
│ ├── agents/
│ ├── scripts-and-code/
│ ├── outputs/
│ └── results-and-learnings/
├── seo-aeo/
├── outbound/
├── paid-and-creative/
└── lifecycle-and-cro/
You don't need to build everything at the start. Start with one vertical, such as content, and create only the content folder.
The folder itself won't tell the Agent how to work. Add a short Markdown file that states where to find things, which sources to use, and where outputs are stored. Each workflow references the files it needs.
Connecting tools and data
The content workflow needs access to X and Typefully. The X connection is used to research posts and performance data, and Typefully is used to create drafts. These connections are implemented through APIs. API keys go in the .env file; don't put them in prompts or knowledge files. Permissions should be separated: creating drafts and publishing are two different things. The workflow should include an approval step.
Shann³ also uses the /last30days skill to search across X, Reddit, YouTube, and the web to understand recent discussions. But remember, skills provide the research process, connections provide the sources, and you still need to check whether the findings are useful.
Building the knowledge layer
Knowledge is divided into internal and external.
Internal: product details, customer calls, CRM notes, and results from previous campaigns. External: competitors, public discussions, and search demand. Shann³ uses a directory structure to store these:
markdown
shared-knowledge/
├── audience.md
├── product-and-offer.md
└── positioning.md
paid-and-creative/knowledge/
├── customer-objections.md
├── approved-claims.md
└── creative-test-learnings.md
Your domain expertise also needs to go into the context. For example, voice notes explaining why the last campaign attracted the wrong buyers, and recording the reasons for feedback during reviews—these can all help next time.
The data layer
If you want to use an Agent to plan the next round of ad creative, the Agent needs to know what happened in the previous round. Data is often scattered across ad platforms, website analytics, and CRM. Aggregate it into a data warehouse, use scripts to fetch it on a schedule, and let the Agent query and analyze it.
Note that API connections don't update data by themselves. Use a cron job to run on a schedule. Records need join fields, such as campaign ID; missing data should be flagged, not blindly attributed.
The first workflow
Pick a task you're familiar with, such as generating the next creative brief from customer research and previous results.
Plan first. Shann³ uses a transcription tool to talk for 5–10 minutes, then uses Matt Pocock's Wayfinder skill (a planning skill originally for engineering, but also suitable for marketing) to refine the plan with the Agent.
After it runs, check the output. If competitor research is irrelevant, fix the selection criteria; if the angles ignore customer objections, check whether the context is missing.
Save the steps as a reusable workflow and keep adjusting it.
Connecting across campaigns
A product launch needs landing pages, paid ads, content, email, and outbound sales. Each vertical has its own workflow, but they all start from the same offer, audience, and plan.
Shann³ uses tickets to track tasks. Each ticket describes the deliverable, references context, and lists checks and dependencies. A coordinating Agent reads the tickets and dispatches them to specialist Agents.
He also uses Paper or Figma to create visual moodboards and a campaign book as references for all channels. When generating images, the campaign book provides product images, lighting, background, and other references to ensure accuracy.
After the campaign ends, results are fed back into the knowledge base for use in the next campaign.
Loops and autonomous tasks
Once a workflow has clear steps and checks, you can let the Agent repeat until it reaches the standard. Define what 'done' looks like, for example:
/goal produce the paid creative set listed in campaign-plan.md.
use campaign-book.md and approved-claims.md.
run the checks in creative-checklist.md.
fix failed checks and save the exports with a review report.
stop when the checks pass and the set is ready for my review.
do not publish. ask if an unresolved decision blocks progress.
Loops need budgets and retry limits, and when they fail, ask why.
Scheduled tasks: daily data updates, weekly competitor reviews, and weekly performance reports, run with cron jobs. Note that scheduled tasks won't make up for an imperfect workflow; test it well before scheduling.
Teams can also assign work to Agents. For example, the Hermes bot in Slack: ask it to "review last week's paid ad performance, compare it with the previous week, and prepare the brief for the next test." The bot needs access to data and context.
Workflows you can build
Shann³ lists some examples: competitor website monitoring, competitor creative library, turning podcasts into team content, SEO production, paid ad creative and testing, content repurposing, email campaign production, landing page production and testing, and creator campaign research.
He is working with @creator_wire on a creator campaign research tool.
Getting started
Pick a task you already have, open a workspace, and build your first workflow. Record where you need to step in; those places tell you what needs improvement. Save the evidence of what you built: briefs, workflows, sample outputs, and results. You don't need to run everything at once; pick one task and start.
发布时间: 2026-11-23 02:01