What I built
The Blog Engine runs across six agents in Claude Cowork, organized in three phases. Each agent was designed and tested as a standalone piece before being connected to the others, which means any one of them can be updated, replaced, or extended without breaking the rest. That modularity was intentional from the start.
The short version: Most AI content tools give you guardrails on voice. That's not the same as understanding perspective. I built a six-agent Blog Engine in Claude Cowork that closes that gap, not by removing the human from the loop, but by making the loop itself useful. The learning compounds because the feedback is captured, not because the AI figured it out on its own.
There's a difference between an AI that knows how you write and one that knows why you said it that way.
The first is a voice guide. Every major AI tool has built-in memory now: Claude knows who you are, ChatGPT remembers prior conversations, and most people have a brand or voice doc of some kind. That gives the AI guardrails: how you want to sound, what you avoid, the register you're going for. It's a real foundation. But it's not the same thing as understanding the perspective behind a particular piece, the specific angle, the reasoning, why this framing and not another. That part lives with you. It's your opinion, your lived experience. You have to share it every time.
That's the gap the system I built is designed to close. Not by removing the human from the loop, by making the loop itself useful.
Phase 1: Research and Brief
The Intake Agent is triggered when I have a new idea. It prompts for the goal, the audience, and what evidence or experience I'm drawing on before anything gets added to the backlog. This isn't admin. An idea that can't answer those three questions isn't ready to develop.
The Triage and Brief Agent turns a backlog entry into a working brief. The brief is the first hard stop. I review it before any draft begins, because structural problems caught here cost one revision to a document. The same feedback caught after a draft exists costs a full rewrite. I learned this the hard way on an earlier post and built the gate in after.
Phase 2: Draft and Polish
Once the brief is approved, the First Draft Agent writes to it. It reads the brief, the voice guidelines, and the existing Empact content that's been published. It's not starting from a generic baseline, it has brand voice loaded and published posts as reference points. What it doesn't have is the inner reasoning behind this specific piece. That's what I provide in the brief, and in the feedback I leave on each draft.
After the draft is complete, there's a second human gate. I review it, leave editorial notes, sometimes dictated, sometimes typed directly into the document, and those notes are what the Polish and AEO Agent reads before it does anything. Every gate requires an explicit signal: approval or direction. That decision is mine to make, and the next phase doesn't start until I make it.
Phase 3: Publish and Distribute
The Publish Prep Agent converts the polished draft into a final version ready to post and generates publish notes. The Visual Direction Agent produces a brief for the post image. LinkedIn repurposing is handled separately by a standalone LinkedIn workflow, not as part of this engine, that's a deliberate separation, not an oversight.
What makes this a system and not a workflow
Every agent produces an output. What makes this a system is what happens at the boundary between outputs.
At the start of each phase, I explicitly ask the agent to read back through the relevant documentation in the folder: the brief, any captured feedback, the ops memory. Some agents have this built into their instructions; for others it's a step I take before triggering the next phase. Either way, it's a concrete action, not a metaphor. It closes the gap between an agent running on thread momentum and one executing against a specific intent.
More importantly, feedback gets captured and fed forward. What I changed in the draft and why. What the agent reached for that I always cut. What's getting closer to my voice without me having to spell it out. That capture happens at the end of each post and becomes part of the context the next one starts with.
This is what I mean when I say the system gets smarter over time. Not autonomously smarter, smarter because the feedback loop is closed deliberately. The AI knew how I wanted to sound from the start. It gets better at the why as the captures accumulate.
Human-in-the-loop is never a limitation, because it's the design. I could click "approve" at every gate and get to a final version faster. I wouldn't learn anything, and neither would the system.
The reference library
Right now the agents reference Empact's voice guidelines and published blog posts. That's enough to keep drafts grounded, but the real power comes when the library gets richer.
Adding to it works like this: I have a conversation with my Claude coworker, share a document, and it goes into a structured folder as an MD file. That file is then available for agents to call on when they need it. It's not a database update. It's closer to filing something in a shared cabinet and knowing where to find it.
The current library is lean: voice guidelines, a handful of published posts, ICP and positioning documentation. The future state is more useful: case studies as they develop, research, client-facing documentation, eventually transcribed conversations and recorded patterns from real prospect interactions. Each addition gives the agents more to work from, and tighter source material means fewer corrections per draft.
The library grows the same way the system learns: through deliberate, documented additions. Not automated accumulation.
The next layer
The Blog Engine is one slice of a larger content system. The LinkedIn workflow runs separately: same principles, different agents, different output format. Right now those two systems operate in parallel but don't coordinate.
The next build is a planning layer that sits on top of both: something that reads the backlog, the LinkedIn idea bank, recent content performance, and what Empact needs to accomplish in a given week, and produces a coordinated plan before anything gets written. That's where the content operation becomes truly connected. The Blog Engine and LinkedIn workflow become modular pieces that a planning layer can direct.
This is still in design. Building toward it has already shaped decisions in the Blog Engine, including keeping the agents modular enough that they can be orchestrated from above rather than just chained together.
What I've figured out from running it
The intake step is the most important part and the one I resist most. When I'm moving fast, I want to skip to the draft. Every time I do, I pay for it in the editing pass. The three questions, goal, audience, evidence, are not process. They're the difference between a piece that says something and one that just sounds like something.
The brief gate is worth defending. Structural changes are cheap before a draft exists and expensive after. Building an explicit human review between the brief and the draft was the single most useful change I made to the system.
Re-orientation is underrated. Asking the agent to re-read the brief and captured feedback before each phase keeps the work aimed at the right target. In a long thread with a lot of history, it matters more than it sounds.
The editing pass is where I figure out what I actually think. There's a difference between the AI getting my voice right and me writing something I'd stand behind. I don't always know which one I'm looking at until I'm in the pass.
Feedback capture is the discipline. Not the most exciting part of the process, but it's the compounding mechanism. Skip it and you're doing the same editorial corrections forever.
What to do with this
Two things separate AI as a one-time accelerant from something that compounds. A structured gate before drafting starts, so the AI writes to a real intent, not an assumption. And a feedback log at the end of each session, captured and fed back in at the start of the next one.
The gate forces clarity. The log closes the loop.
Structuring business context so AI has something real to work from, source material, voice, positioning, offer clarity, is the Foundation engagement at Empact. Start here to understand how it applies to your situation.