Two days before, my morning briefing arrived in the format it had been taking all week. A plain text summary in my inbox at seven in the morning, generated by an automation I built in Claude Code earlier this month. The part I want you to see, verbatim, looked like this:
📤 PUBLISHING TODAY
No blog post. Social drafts are staged and ready in
Social/social-posts/2026-04-16/— anchor is Why Is My Hair Curly? (Week C/Thursday). The 9 AM “BTS OVERRIDE” calendar event is a misfire from yesterday; dismiss it. Fill the[PLACEHOLDER]in Instagram + Facebook (optional book scene), swap[link]for the Amazon URL, and post.
I read the briefing. I filled in the placeholders. I posted the drafts. Fifteen minutes, start to finish.
On Monday, the same pipeline ran for the first time with a real blog post attached. The essay was called What You Take Out. It was about the art and craft of revision, the grunt work of shedding what is needed to make what remains shine. It went live at seven in the morning. When the briefing hit my inbox, it told me the post was published, pointed me at a folder where social cards had been waiting since Sunday night, and listed the platforms I needed to publish to by hand. The cards carried an image that nano-banana, an image generation model, had produced on Sunday from a prompt my automation wrote after reading the essay itself. The image was branded in the colors I use across everything. Burnt sienna. Charcoal. Cream. The Canva templates had been built for those colors weeks ago. The blurbs for each platform were already written, because the automation drafts them from the essay, borrowing my words.
I sat at my counter with my coffee. I read the briefing. I read the drafts. I fixed whatever needed fixing. I published.
It replaced something that used to take me two hours.
Let me walk you through what the two hours looked like.
Before any of this was built, here is what happened when I wrote an essay for my blog. I would finish the draft. I would open a new tab and go to Midjourney or search through free royalty free photo sites. I would write a prompt describing an image that might represent the essay. I would wait for the generation. I would usually reject the first four images and try again. I would eventually get something close. I would download it. I would open Canva, or more often a free online cropper, and make one version for the blog share, one for Instagram at a square ratio, one for Pinterest at a tall ratio, one for Facebook at whatever the current Facebook sizing was. I would upload the blog version to WordPress and assign it as the featured image. I would open Canva and drop the other versions into my branded templates, the burnt sienna and the charcoal and the cream, and add the title overlay and my handle. I would export each one. I would upload them to a folder. I would open Instagram. I would write a caption. I would post. I would open Facebook. I would write a different caption. I would post. I would open Pinterest. I would write a third caption with the keywords I cared about that week. I would post. I would open Threads. I would write something shorter. I would post. I would open Substack Notes. I would rewrite it for the Substack crowd. I would post. I would open X, which I no longer use much but still cross-post to. I would shorten it further. I would post.
This took two hours on a good day. On a day when I am not able to find the right picture for my post, or when the sizing broke in Canva, or when one of the platforms had changed its interface overnight, it took three.
I did this work on top of the writing. I did it on top of the day job. I did it on top of the three daughters, the marriage, amma waiting to hear what I made for lunch, the manuscript that is currently under revision, the novel in draft, the two kidlit books under contract, and the existence of being fifty years old and responsible for a household.
This is the grunt work of being a writer with a platform in 2026. No one talks about it honestly because every hour of it is an hour the writing did not get. Writers are supposed to pretend the writing is the thing, which it is, and that the platform is something that happens on the side, which it is not. The platform eats the hours the writing needs. The writing community I am in knows this and does not talk about it, because acknowledging it means acknowledging the bind.
Here is the bind. A writer who does not maintain the platform will not be read. A writer who does maintain the platform, at the level the algorithms now demand, will write less. The time has to come from somewhere, and it comes from the writing.
I spent most of March and early April trying to get out of that bind without saying that was what I was doing.
Most of March, I tried to build my way out of the two hours using Cowork.
Cowork was the second interface I learned about, after the chat window. It had scheduled tasks. It had project folders. It seemed, from where I was standing, like the setup that would let me automate the image pipeline I needed. I spent weeks on it.
I got close. I built an automation that could trawl the news and the internet for post ideas and tell me what was happening in the world that applied to me, a writer reflecting on a hyphenated existence. I built another that could fetch the hero image from a folder I dropped it into and size it into Canva templates. What I could not do was close the loop. Every day the automation would give me prompts. Every day I would open a separate tab, go to Midjourney, get the image, drop it in the folder, let Cowork pick it up. The pipeline was faster than doing it all by hand, but the pipeline still required me to stop whatever I was doing, mid-morning, and feed it an image.
I kept looking for a way to close that gap. I wanted the image to come from the same place the post came from. I wanted an image model I could call without switching tabs.
I found one. Nano-banana. A tool for generating images through the Gemini API. There was a way to connect it to Claude. I got the API key. I set an environment variable.
I spent seventeen years writing software. I left that work in 2014. In all those years, I worked inside IDEs, the graphical environments where code is visible and structured and the interface is designed for humans who are writing, not for machines that are running. I did not love the command line. I loved the part of software that felt like composition. When I came back briefly to development in 2022, I confirmed what I had known in 2014. The grunt work of coding still bored me. The command line still made my jaw tight. I did not want that life.
So when I read, in March of this year, that setting up nano-banana would require me to get an API key from somewhere, set it as an environment variable, and then configure something called an MCP server, I understood roughly what each of those words meant. I also did not want to do any of it. The instructions were written for people who lived at the command line. I did not.
But I wanted the image pipeline.
I tried anyway. I got the key. I set the variable. I wired up what I thought was the connection. Cowork refused to see it. I researched. I spent hours at it, in the cracks between dinners and mornings and the manuscript that was still in revision and amma asking what was for dinner and the platform post I still had to write for that day. Every path I tried, Cowork’s sandbox blocked. I did not understand at first that the limitation was structural. I thought it was me. I thought I was setting something up wrong.
I filed a feature request with Anthropic. I asked them to add nano-banana support to Cowork. I do not remember exactly when I sent it or whether anyone answered. The request itself is not the point of this story. The point is that filing it was the move of a woman who had given up on making the integration work herself. I was asking someone else to build the door I could not find.
By the time I gave up on Cowork, I had decided the image pipeline was not going to happen. Not this month. Maybe not this year. I had hit the edge of what my tools could do, and I did not know how to get past it.
Then, about a week ago, I did something I had not done since 2014.
I installed Claude Code.
The reason I installed Claude Code was not curiosity. It was the morning briefing. The automation I had built in Cowork was generating a summary of my inbox every morning, listing items I needed to act on. It was useful. It was also circular. The same items showed up every day because I had no way to tell the system I had handled one. The briefing said pay tuition fees, I paid offline, there was no online record of me paying the money in a way Claude could see, the briefing said pay tuition fees the next morning. I needed a way to close the loop.
The small problem I was actually trying to solve, when I sat down at the terminal that Saturday morning, was task tracking.
Google Tasks had what I needed. An API. A way to mark things done on my phone and have them disappear from tomorrow’s briefing. The existing integrations did not connect it to Claude. Someone with more patience than I had could probably have figured out a workaround in Cowork. I did not have more patience. I had already spent it on nano-banana.
In Claude Code, I built the Google Tasks connection in an hour. I do not know how else to describe it. I typed what I wanted. I followed what Claude told me to do. The OAuth flow ran on my phone. The tasks appeared. I checked one off. The next morning’s briefing skipped it.
That was the moment.
It was not a large moment. It was one small problem solved cleanly. No sandbox. No feature request. A task list that updated when I asked it to. But something in me recalibrated. If the command line had been like this twenty years ago, I might not have left.
I want to be careful with that sentence because it is not fully true. The command line was not like this twenty years ago. I have not been wrong for the last twelve years to prefer writing to coding. What is true is that the command line in April 2026, with Claude sitting in it, is not the command line I left. It is something else. A place where a person who fears UNIX can ask for what she wants in English and watch the English become a working integration.
I went back to nano-banana that afternoon. Not in Cowork. In Code. I got the same key I had gotten in March, out of the same environment variable. The MCP server came up in fifteen minutes. I hit two small problems. Something called npx was not where Claude Desktop could see it. I installed Node.js. That took another fifteen minutes. Then the model name in the package was slightly out of date because the model had been renamed. We patched it in place. Then the API key was not enabled for image generation on the free tier. I turned on billing, a few cents per image. I tried a generation.
It worked.
The first image was abstract: a charcoal background, a sienna circle, cream threads radiating. An art-installation image. Fine. The second image, which I asked for after giving Claude the actual blog post I was working on, was a brass tumbler with steam rising. Warm kitchen light. A saucer. It looked like the essay. It looked like something I would have spent an hour trying to find on a stock photo site and not found.
I sat at my desk for a minute before I did anything else.
The pipeline I had been trying to build for five weeks was working. The sandbox was behind me. The feature request did not matter anymore. What had changed was not the difficulty of the problem. The problem had always been solvable. What had changed was that I was in a room where the problem could be solved.
The pipeline ran on a Monday morning. It replaced two hours of platform work with fifteen minutes of review.
What it does is give me back the hours. Two hours on Monday morning that used to go to image sizing and caption rewriting. Two hours that now go to whatever the writing needs, or to the girls, or to amma’s question about dinner, or to the manuscript that is still in revision, or to sitting quietly with nothing to do, which is sometimes the hour I need most.
I sat at the counter with my coffee. I read the briefing. I fixed what needed fixing. I published.
Then I went to write.
The AI Journey
- The Morning I Downloaded VS Code — The morning I opened the terminal I had walked away from twice.
- Two Hours on a Monday Morning — A publishing pipeline, five weeks of failure, and the fifteen-minute replacement.
- Nine Point Six Percent — The SEO audit that asked me what twenty years of writing were for.
- The Librarian and the Historian — Reading across fifteen years of correspondence for a memoir.
- What the Dashboard Doesn’t Count — The cost of the migration that went wrong.
- Saturday Morning — The practice I have concluded for myself.
Leave a Reply