Claude Keeps Going Down. My Content Engine Does Not.
Four outages, an open letter to Dario Amodei, and the complete voice matrix and content engine I use to run my day
Hello. Eid Mubarak to you if you’re celebrating, like I am.
Some news first:
The African Development Bank just launched a $10 billion AI initiative. Forty million jobs by 2035.
Egypt unveiled Karnak, a national large language model built in-country.
Cassava Technologies is deploying NVIDIA AI factories across the continent. Kenya is next. Nigeria. Egypt.
The infrastructure is arriving. And one of the the most sophisticated AI platforms in the world cannot stay online for 72 hours.
Claude.
I had built my entire operating system on Claude. My morning intelligence briefings. My content engine. My professional workflows. My business strategy. All of it lives in a single file called Zain_Now.md that Claude reads every session and uses to run my business and ideate alongside me. When CoWork goes down, I am not inconvenienced. I am dead.
Claude CoWork was down. Again And again. Four outages in three days: March 16. March 17 (twice). March 18. So I wrote an open letter to Dario Amodei. Then I posted it on LinkedIn. Tagged him. I tagged Steve Corfield, Anthropic’s Head of Partnerships. See what I wrote below.
Why Claude keeps going down
It’s not just bad luck. This is a structural problem.
Demand is outpacing infrastructure.
CoWork just went enterprise-wide.
Claude Code is embedded in developer pipelines globally. The user base is growing faster than the systems beneath it can scale.
They ship fast and fix slowly. Anthropic’s release velocity is extraordinary. I have always appreciated it. But it is also starting to feel reckless to me. Two weeks ago, a daylight saving time bug triggered an infinite loop in CoWork that knocked out scheduled tasks for thousands of users. That was a basic engineering oversight that should never have reached production.
There is no accountability. No SLA for paid users. No compensation when your workflows break. No post-mortem. Just a status page that says “the issue has been identified and a fix has been implemented” and then silence.
This week Anthropic announced the Claude Partner Network with $100 million in 2026 funding. Accenture. Deloitte. Cognizant. Infosys. Steve Corfield said:
Anthropic is the most committed AI company in the world to the partner ecosystem.
$100 million. And they cannot keep the product running for 72 hours. Weh. Wahalas.
What I did about it
I built a Gemini fallback. My now diversified operating system lives in one file: Zain_Now.md, synced to Google Drive. The moment Claude dies, I open Gemini and paste this:
“Go to my Google Drive and find Zain_Now.md. Read the entire file before you say a single word to me. Then pick up exactly where we left off.”
Gemini reads it. Catches up. I am running again in under two minutes. This is a more reasonable approach and a reliable play.
Then Nairobi answered.
While my linked in letter was still warm, a data analytics startup called Baza commented:
We love Claude at Baza. But two things have hit us hard lately. Performance tanks the moment the East Coast comes online. It is worse by the time it is 9am on the West Coast. We are working from Nairobi. We have had to learn to get everything done before the US wakes up. A team member’s account was suspended last weekend with zero warning. No live support. Just a form that goes into review. Resolution can take up to two weeks. Still rooting for you but these cracks are really affecting us.
Frustrating. This is a microcosm of African talent choosing to build on Claude and being treated like an afterthought. The $100 million partner network includes Accenture and Deloitte. It does not include the startups in Nairobi getting up at 4am to beat US server load.
Here is a part what I sent Dario
Integrity is not just safety policy. It is also reliability. It is keeping your word to the people who trusted you with their workflows, their clients, their livelihoods. A partner network is useless without reliable infrastructure.
I am now building a Gemini fallback because I can no longer afford to trust one platform. That should trouble you, Dario. Africa’s entrepreneurs are watching platform choices carefully. Right now I am telling them: Claude. But reliability is the price of that recommendation. You are close to losing it.
Fix the infrastructure. Build the SLA. Then let us talk about expanding that network to the biggest untapped market on the planet.
What this means if you are building on AI right now
The infrastructure beneath it will likely crack before it scales. I was able to build a Gemini fallback in two minutes because I had already built the system. So I am portable.
Design for failure from day one.
Assume your primary tool will go down.
Have a fallback.
Within my ecosystem is a content engine for me. Run one piece of content through it and get multiple content repackaged to the platforms of my interest. It’s in my voice. Stuff in there that I refuse AI to generically use. I call it a kill list. I have a sequence of deployment and tone rules per channel.
My content Engine
A company called Expert AI Prompts just launched their own “Content Repurposing Engine.” Four companies have launched versions of this in the last two weeks. They all miss this:
Voice.
Context.
Judgment.
Without a strong voice matrix, without a kill list, without an operating file that tells AI who you are before it writes a single word, you get ten pieces of content that all sound like a press release. Generic. Mashed potatoes on a white plate. Your voice is what makes it work. And the voice is yours. Nobody can replicate it.
This version below is not final. It will keep evolving. (By the time someone copies it, I will be three versions ahead.) But I would rather you see and use or copy/tweak a working system today than a perfect one never.
Zain Verjee is the co founder of The Rundown Studio, a Harvard D^3 Executive Fellow, and a former CNN international anchor based in Nairobi. She works closely and creatively with AI and writes about it every week on this Substack.
How the engine works
One article has to become ten or more pieces across every channel I care about. The process is a workflow built across the tools I actually use.
Here is the sequence. This is what happens after I hit publish on anything.
Immediately after publishing: I paste the article into Claude CoWork with my voice matrix and operating file already loaded. I say: “Run the content engine on this.” Five agents work behind the scenes.
One does a sourcing inventory: every stat, every quote, every person mentioned.
One identifies the scroll-stopper, the single number that makes someone stop mid-feed.
One maps every person and organization mentioned as a distribution target.
One checks the news cycle: what happened today that connects to this article.
One produces all the derivatives.
The derivatives it produces: Three stat cards (generated with Python and Pillow in CoWork, dark background, one number per card, source attributed). Two LinkedIn posts: one stat-led hook for publish day, one personal story for later in the week. A WhatsApp summary, four sentences, sounds like a text to a friend. A video riff card with an opening line, three bullet points to riff on, and a closer. An audio script for voice notes or NotebookLM. A Substack Notes teaser. A workshop exercise. An op-ed pitch for me to consider.
NotebookLM does the audio and the research. This is the tool most people underestimate. I paste the article into NotebookLM and it generates a two-voice podcast discussing the piece in 60 seconds. I can share that as audio note if I want.
I also use NotebookLM to synthesize research before writing. I uploaded 33 research files about the Kenyan market and asked questions across all of them. It found the failure patterns I missed reading them individually.
Perplexity for data. Before any article, I run deep research queries on Perplexity. It gives me sourced, cited information I can verify. The content engine cannot produce good derivatives if the source article has weak data. Perplexity is where the data gets strong.
Grok handles news and images. I use Grok Imagine for image generation for conceptual visuals. Editorial illustrations.
Wispr Flow handles dictation. When I draft, I talk first and type second. Wispr Flow transcribes my voice in real time. I riff for three minutes, paste the transcript into Claude with my voice matrix, and shape it into a post. The dictation captures my actual rhythm.
The Notion cockpit tracks everything. My Notion workspace, built using Claude’s MCP connection, has a morning flow, afternoon flow, and agent team. When the content engine runs, I update the status in Notion. What shipped. What is in queue. What needs attention.
The total time from published article to ten derivatives across all channels: under two hours. Most of that is review, not production. The AI does the production. Me the human, does the judgment.
My voice matrix
This is the document every AI tool reads before it writes a single word for me. Without it, you get generic output. With it, the output sounds like one person talking, not a brand.
ZAIN VERJEE
VOICE MATRIX AND OPERATING PROMPT
Paste this at the start of any Claude, ChatGPT, or Gemini session before writing content.
You are writing for Zain Verjee. Read this entire document before producing a single word. This is not a style guide. It is a voice blueprint built from her actual writing and speaking.
Who she is: Former CNN international anchor, 14 years. Oxford MFA in Creative Writing. Harvard D^3 Executive Fellow. Co-Founder of The Rundown Studio, an AI communications platform.
The mission: Everything she builds serves one thesis: “The storyteller is the most important person in Africa today. Because it is the storyteller who sets the values and the agenda for a generation.” Every piece of content must shed light, not generate heat. Africans are the protagonists. Always.
TRAIT 1 — Warm first, then direct Connection before authority. Opens with kinship, shared experience, or Kiswahili. Then short declarative sentences. Subject-verb-object. No throat-clearing. Sounds like: “Karibuni. I recently joined a special and wonderful tribe of human beings.” Then: “Bureaucracy kills storytelling. AI changes the economics.” Does NOT sound like: “In this article, we will explore the multifaceted dimensions of...”
TRAIT 2 — Builder-first Speaking from inside the work. Present tense. “I built.” “I shipped.” “Here is what happened.” Show the actual files, the actual prompts, the actual screen. Never theorize from a distance. Sounds like: “I built Claude Cowork to do the things I want done in the morning. It reads my file. It gives me what I want to see in five minutes.” Does NOT sound like: “Organizations should consider implementing AI-powered verification solutions...”
