Shann Holmberg: AI Marketing Engineer Workflow

Main Tweet (2026-05-03)

this is how you become an AI Marketing Engineer

1. prototype the workflow in hermes agent or openclaw.

2. run it two or three times against real work, correct it when it drifts, let the harness write the skill on the way.

3. once the shape is right you pull it into a dedicated workspace in claude code and tighten the execution, the prompts and the routing, the cron and the failure modes.

4. when it survives a week of real runs without you babysitting, you push it to a VPS and let it run on a schedule

prototype in hermes agent or openclaw fine tune in a dedicated claude code workspace deploy autonomous on a VPS

Stats: 179 likes, 17 retweets, 19 replies

Quoted Thread (2026-05-01, 8 tweets)

1/8

how to prototype marketing workflows in Hermes Agent run the workflow two or three times against real work, correct it when it drifts, the agent watches each run and writes the skill along the way once its solid, you take it into claude code for the detailed execution or automation

2/8

I map the workflow inside hermes, run it two or three times against real work, correct it when it drifts, let the harness watch each run. by the third pass it knows what the first two taught it

3/8

I dont run marketing through one model, each job goes to the one thats actually good at it

Opus 4.7 for copywriting GPT 5.5 for planning, strategy, sometimes coding GPT image model 2 for visuals Codex 5.3 or Kimi 2.6 when a coding step lands inside hermes you wire those into the same workflow, each step routes to the model thats good at it

4/8

the harness writes the skill itself. capture triggers on task completion past five tool calls, on error recovery, on your corrections, on a workflow it hasnt seen before. the artifact gets written during the run, and the next run picks up where the last one ended

5/8

memory persists across sessions in ~/.hermes/memories/, MEMORY.md and USER.md inject at session start, sessions stored in sqlite. your voice rubric, brand notes, customer language, last weeks corrections, all of it loads before the first prompt. no monday retraining

6/8

v0.12 shipped a curator. background agent on a 7-day cron that grades the library, consolidates skills that overlap, prunes the dead ones. the skill folder doesnt rot the way it usually does when nobody touches it for a month

7/8

once a workflow is solid in hermes, I take it out. either pull the skill into claude code and run the detailed execution there, or wire it to a cron inside hermes if it runs on a schedule. prototyping in hermes, execution in claude code, thats the split that works for me

8/8

next thing im testing in hermes is the brand voice rubric as a curator-graded skill, see if it survives a month of corrections without me babysitting it

Notable Replies

• "Can I set it up on vps?" → "Yes, that's the best way to set it up"
• "And do you sell this to clients?" → "Yes either that or you run a business and want to automate"
• "The feedback loop is pure alpha. Human in the loop keeps the agents from drifting"
• "third run it clicks huh"