My Substack used to take 12 to 15 hours a week. Research, drafting, editing, formatting, scheduling notes, responding to readers. None of it was creating. All of it was operating.
Now it takes about 2. The research runs while I sleep. The draft appears in my Substack editor by morning. Notes go out on a schedule. Engagement gets surfaced so I can respond fast. One system handles all of it.
This is what I built, how it works, and how you can start using the same foundation today.
The Workflow That Was Killing My Output
Every week looked the same. Pick a topic. Open 15 tabs. Read for two hours. Synthesize. Draft. Edit. Format. Schedule notes. Space them manually. Check comments. Respond.
The writing was never the hard part. The operating was what took all day.
The deeper problem: every time I opened a new AI session, it started from zero. It did not know what I had already written. It did not know my voice, my recurring topics, or the direction I was taking the publication. I spent the first 10 minutes of every session re-explaining myself. Then I re-explained it again next week.
The cost compounds. Not just the minutes. The context switching. The cognitive load of holding in your head what the agent has already forgotten.
What the System Does Now
When I start a new article, I run one command. The system gathers intelligence on the topic across multiple sources, synthesizes what is worth saying, writes a draft in my voice, generates illustration prompts, produces the Substack draft with correct formatting and tags, and queues five scheduled notes timed for engagement.
Seven phases. Zero copy-paste between sessions. No re-explaining what my publication is.
That is a single article mission. Each phase hands off to the next with full context preserved. The researcher passes findings to the writer. The writer passes the draft to the note strategist. The note strategist passes to the scheduler. By the end, everything is done.
The system also handles engagement. It scans Substack, LinkedIn, and Reddit for relevant conversations, drafts responses in my voice, and surfaces them for approval. I spend 20 minutes instead of two hours.
200 missions. 91 percent success rate. 25-minute average from task to done. This is not a demo environment. This is the production system that wrote this article, generated the illustrations, and created the Substack draft you are reading right now.
Why Context Is the Entire Game
The research backing this is worth knowing in brief. Vercel benchmarked AI agents with and without a static project context file. No context: 53 percent task success. With a well-structured CLAUDE.md file: 100 percent. Boris Cherny, who built Claude Code at Anthropic, runs the same pattern: every mistake the agent makes goes into the file so it cannot make that mistake again. Google's context engineering whitepaper frames the outcome plainly: "Agent without memory = tool. Agent with memory = colleague."
The system that automated my Substack workflow is built on this principle. The agent does not relearn my voice, my topics, or my conventions each session. That knowledge persists. And it compounds.
What the Starter Kit Includes
The foundation of the system is available now. The Via AI Starter Kit gives you the context layer: the CLAUDE.md rules, the persona files, and the memory architecture that make all of this work.
CLAUDE.md with five production-tested rules. Not a generic prompt template. Five specific behavioral corrections from 200 real missions: read existing work before modifying it, only make requested changes, avoid premature abstractions, do not create helpers for one-time tasks. These are the rules that survived production.
Three personas. A writer for drafting. A researcher for gathering intel. A reviewer for quality checks. Each loads automatically based on what you ask. Say "research" and the researcher activates. Say "write" and the writer takes over. No configuration required.
A memory guide and MEMORY.md template. The three-tier memory architecture translated into a practical setup. Semantic memory in MEMORY.md, loaded every turn. Episodic memory in project-specific files. Procedural memory in CLAUDE.md itself. No vector database. Just markdown files that fire at the right moments.
A skills routing index. The compressed format that tells the agent what tools exist and when to use them, without loading full documentation into every turn.
The install is one step. Download the kit from joeyhipolito.gumroad.com/l/ai-starter-kit, unzip it into your project directory, and start your first session. The agent reads CLAUDE.md automatically on every turn.
After three sessions, the file reflects your publication rather than a template. That is when the compounding starts.
What the Full System Adds
The Starter Kit is the context layer. The full Via system adds orchestration on top of it.
Instead of one agent running one task, the orchestrator chains 21 specialized personas through multi-phase missions. The researcher gathers. The writer drafts. The cartographer maps diagrams. The artist generates illustrations. The note strategist queues engagement. Each phase runs in sequence with full context passed forward.
The result is what the dashboard shows: 200 missions completed, 91 percent success rate, a publication that runs on a fraction of the manual work it used to require.
Start Here
The free Starter Kit is available now at joeyhipolito.gumroad.com/l/ai-starter-kit.
Download it. Run your first session. Then open CLAUDE.md and add one thing you already know about your publication that the agent does not. That first manual edit is the most important step. The template gives you the structure. The edit gives it the specific knowledge that turns a generic starting point into something that actually reflects how you write.
If you want to see what the full system looks like for your Substack workflow, reply to this post or reach out directly. I am building the paid tiers based on what creators actually need, and the most useful thing you can tell me right now is which part of your workflow costs you the most time.
What would you automate first?