Four months ago, I sat down and mapped my entire content pipeline on one page. Research in. Digest out. Digest feeds a draft. Draft splits into four platform posts.
That session took three hours. It’s the last time I had to actively think about the structure.
Every week since, the pipeline runs: one long-form article, four social posts across LinkedIn, X, and Substack Notes, two email brief drafts, one research digest. I spend about 3 hours on all of it. Three Claude agents handle the rest.
Before the system: 8 to 10 hours a week. Research scattered across sessions. Reformatting the same ideas for LinkedIn, then the email version, then Substack. Different mental states, different quality levels, no consistency.
After: 2 to 3 hours. Same output volume. More consistent quality — not because I got better, but because the structure doesn’t change.
Why One-at-a-Time Prompting Doesn’t Scale
The instinct when you first use Claude seriously is to treat it like a smarter search engine. One prompt, one output, done.
It works fine for a single task. It breaks down for anything with multiple steps and dependencies.
Content creation is dependencies all the way down. The research shapes the draft. The draft determines which angles work for social. The social posts need to feel like they’re from the same piece, not four different writers. If you’re prompting each stage manually — copy-pasting between sessions, re-establishing context each time — you’re spending half your effort on connective tissue instead of the actual work.
The agent architecture solves this by making the connective tissue automatic.
The 3-Agent System
Three agents, fixed sequence, no manual handoff:
Agent 1: Research. Takes a topic or keyword list. Calls search tools, pulls RSS feeds, fetches specific URLs I’ve bookmarked. Outputs a structured digest: key claims, notable quotes, gaps in the current conversation, angles that haven’t been covered well.
Agent 2: Drafting. Takes the digest. Produces a long-form draft in my voice — direct, specific, no AI-sounding filler. The draft includes callouts for which sections might work as standalone social posts.
Agent 3: Reformatter. Takes the draft. Produces four platform-specific posts (LinkedIn hook + body, X thread, Substack Note, and an email brief). Different format, same core idea, different editorial decisions for each platform’s reader.
Each agent has a fixed system prompt. No improvising. No “write a post about X” prompts that produce different quality each time.
What Each Agent Actually Needs
The agents work because their inputs are constrained, not because the model is smart enough to figure out what you want.
Agent 1 needs: a topic brief (2–3 sentences max), a list of source types to prioritize, and a clear output format for the digest. If the output format isn’t fixed, Agent 2 gets inconsistent input and the draft quality swings.
Agent 2 needs: the full digest from Agent 1, a voice reference (I use a file in Claude Projects with 10 examples of writing I want to match), and word count + structure constraints. Without the voice reference, it defaults to generic AI prose — you’ll know it when you see it.
Agent 3 needs: the full draft, a character-count spec per platform, and explicit instructions about what NOT to do (no em dashes on LinkedIn, no clickbait hooks on Substack, no passive voice anywhere). Negative constraints do more work than positive ones here.
Real Weekly Numbers
Four months of data, same pipeline:
Average research time (my active involvement): 45 minutes, reviewing and pruning what Agent 1 pulled.
Average drafting review: 40 minutes, editing Agent 2’s output.
Average reformatting review: 30 minutes, tweaking Agent 3’s platform posts.
Total: around 2 hours 15 minutes of active work per week. The rest is the agents running.
Before: I was spending those 8–10 hours doing what the agents now do, plus the review time. The review time didn’t go up — the production time went to zero.
What to Build First
If you’re starting from zero, build Agent 1 before anything else.
The research agent is where most of the value is. Drafting is the glamorous part, but bad research produces good-sounding bad drafts — and you won’t catch it in review until the piece is already half-done.
Get Agent 1 producing a structured digest you’d actually use. Test it on 10 topics. If the digest is good, Agent 2 is straightforward — you’re just giving a competent model solid input.
The reformatter (Agent 3) is the easiest to build and the lowest value. Don’t start there.
The System Is Already Built — You Just Need to Run It
The architecture described here isn’t a framework or a concept. It’s the actual system I run every week.
The piece you’re reading came through it.
What makes it work isn’t sophisticated AI. It’s fixed structure and disciplined constraints. The agents do what they’re told, consistently, because they’re told specific things in the same format every time.
Variable prompts produce variable output. Fixed systems produce reliable output. That’s the entire lesson.
MCP Explained 2026 · AI Agents Decoded 2026 · Stop Explaining Yourself to AI · Complete Guide to Claude Chat, Cowork, Code