How my content marketing workflow became fully agentic in a year

Almost a year ago I published Why Startups Don't Need AI Agents (Yet). The argument held up for its moment: for a solo founder or a small team, custom GPTs, prompt libraries, and scheduled scrapes covered most of what people were reaching for agents to do, at a fraction of the cost and fuss. But the whole thing rested on one small word at the end, and that word was doing a lot of work. The "yet" has arrived. A year on, my content marketing runs on agents from one end to the other, and I want to write down what that shift felt like from the inside.

What a day looked like a year ago

When I wrote that post, my day was manual in a way that is almost funny to describe now. I opened ChatGPT and bounced between it and Claude depending on the task. If I wanted to build something, the loop was ask for code, copy it out, paste it in, run it, watch it break, paste the error back, read the explanation, edit by hand, try again. I was doing the orchestration myself, with a lot of copy and paste in between. At the same time I was circling the tools everyone was circling: n8n, Cassidy AI, Relevance AI, Make. I kept looking for the one that would tie my workflow together, and none of them quite fit the shape of the work I do.

The split that made it click

What changed is not that I found the magic tool. It is that the models got good enough to do the orchestration themselves, and I settled into a division of labor between two of them. These days I pay for two tools and spend most of my time in Claude, which does the building. I do not like asking Claude to check its own homework, so ChatGPT does the reviews, and the notes go back to Claude to fix.

ChatGPT also plays a second role, closer to a sparring partner. When I have a half-formed idea, it is the one I think out loud with, because it holds context and memory well and can pull something I said weeks ago without me re-explaining. That matters more than it sounds. With Claude, the memory I care about is usually tied to a repo, which means I am maintaining a second brain: keeping it current, not drowning it in files until it gets worse at its job, following all the small context-engineering habits that keep it useful. That upkeep is work, and after a while it gets tiring. So I lean on the tool that remembers on its own for the thinking, and the tool with the repo for the building.

Marketing work, in the terminal

The part that surprised me most is that a lot of this now happens in the terminal. A year ago the idea of doing marketing work there would have made me laugh, and I was skeptical that Claude Code had anything to offer someone whose job is content, not software. Now it is a daily tool, sitting next to the Claude and ChatGPT chat windows, and I reach for whichever one fits the task in front of me.

MCPs, the boring magic

Back in January I wrote a post working out what an MCP even was, in slightly saltier language than I will use here. Now I use them for small things constantly, with one firm rule: I do not let AI delete. I do not trust it to rummage through my things and decide what to remove. Adding and rearranging, though, I hand over gladly.

A few that earn their place:

• Asana. When a webinar slipped by two months, every task tied to it needed new dates. That is the kind of tedious admin that used to mean me guessing deadlines one by one. Instead I connected the Asana MCP to Claude and asked it to shift everything to fit the new date. It is a small thing, and it made my week.
• Calendars. If I am reading a post with a list of deadlines in it, I drop the text into a chat and ask it to fill my calendar with the dates.
• Site audits. My site lives in a Git repo, and I used to be a little skeptical of people who kept their sites on GitHub or GitLab. Those people are the ones having the best time right now. Connect the repo, run an audit, and walk through the fixes one at a time: meta titles, descriptions, alt text. Claude Code makes the edits while I watch that it is not doing anything strange.
• Social. My social calendar is an Asana board too, so the same connection can draft a week of content straight into tasks. I review, adjust, and design the graphics.

Podcast and video

For the podcast, the editing is light: pulling filler, trimming the odd sentence, cleaning up subtitles in Riverside. I have an agent review the whole thing and hand me back a corrected SRT to upload, and another one drafts the YouTube description with a few title options to choose from. I turned that flow into a Claude Code skill. Years ago I got paid to review subtitle files by hand, and it was miserable. That whole category of work is mostly gone for me now.

Claude's design feature has come a long way as well. Give it a set of design guidelines and it does a decent job on graphics, and you can export a carousel to Canva to finish it off. It is heavier on resources than I would like, so I still do most quick graphics in Canva directly, but it is a nice option to have when I want it.

Where my time goes now

Here is the change that matters. A year ago the work was: have the idea, research the shape of a good landing page, design every part of it, build it, then maintain and optimize. The middle of that, the building, took the most time and gave the least back. Now I bring the idea and the outline, hand it over for a first wireframe, and tweak every part until it is what I want. The implementation step got much lighter.

So my time has moved. It sits in ideation and research at the front, and optimization at the end. What makes a good landing page, how to structure it, how to tune it for the way people search now, including light AEO and GEO work. Those are judgment calls, and they stay mine. The building in the middle is the part I now share with the models.

None of this runs itself. You set up guardrails, you mind the security pieces, and you keep it away from anything you would not want undone. But the shape of the day is different: less grinding through implementation, more time on what to make and whether it is worth making.

That one word, "yet", was carrying the whole argument a year ago, and I knew it when I wrote it. Startups did not need agents then. A lot of them do now, or will soon, and the tools finally fit the work. If you read the original post, read it as a snapshot of a moment that has already passed. This is where I am a year later.