I Run a One-Person Newsroom That Ships 4 Pieces a Week. Here's the 5-Agent Claude Setup.

A walkthrough of the editorial pipeline I built to publish Medium articles, LinkedIn posts, YouTube scripts, and Twitter threads solo. The five Claude agents, the handoffs, the weekly cadence. A chapter excerpt from the book.

The wrong way to publish solo

When I started building a real publishing cadence around my indie work, I did it the way most solo creators do. I'd sit down on a Sunday with the intention of writing "some content for the week," I'd open a fresh document, I'd stare at it for forty minutes, I'd produce one half-decent draft, and I'd convince myself that was enough.

This produces, on average, one piece a week. Sometimes two if you're lucky. It produces nothing the week you're sick, traveling, or distracted by an actual product launch. It produces nothing in any seasonality dip. It produces a wildly inconsistent voice because the one-shot drafting model is reactive to your mood that morning.

I now run a one-person newsroom that ships four pieces a week, every week, across Medium, LinkedIn, YouTube scripts, and Twitter threads. It is built around five specialized Claude agents and a weekly rhythm. The total time I personally spend on the editorial side is about four hours a week. The rest is the agents doing the work that, in a real newsroom, a small team of specialists would do.

This is the chapter from Part 4 of my book — How to Claude — The Playbook for a One-Person Business — that walks through the newsroom build end to end. The chapter is called "The one-person newsroom." Here is the version with all five agents specified.

The five-agent org chart

The newsroom has five Claude-driven roles plus me as orchestrator. Each agent has a Skill file that defines its workflow. Each agent has a CLAUDE.md that defines its voice. The handoffs between agents are explicit artifacts.

+-------+   topics   +--------+   angles   +-------------+
| ANDY  | ---------> | RESCH  | ---------> | ANGLE FINDR |
+-------+            +--------+            +------+------+
   ^                                              | shortlist
   | chose                                        v
   +<------------------------------------- ANDY (chooses 1)
                                                  |
                                                  | brief
                                                  v
+--------+   pub'd    +----------+   final   +--------+
| PUBLR  | <--------- |  EDITOR  | <-------- | DRAFTR |
+--------+            +----------+           +--------+

Five Claude roles. Each one runs as a separate Claude session with its own Skill loaded. Five sessions sounds like a lot until you realize each session does one thing, in scope, in maybe 15-30 minutes of compute time per piece.

Role 1: Researcher (45 min per cycle, weekly)

The Researcher's job is to surface what's happening in my domains every week. It runs once on Sunday evening. It reads my Pinecone memory layer, my Obsidian sessions vault from the past week, and a hand-curated list of news sources (Hacker News digest, indie-dev Twitter highlights, App Store category top changes, my own RSS reader's unread queue).

It produces one artifact: a markdown file called weekly-research-{date}.md with three sections — what happened in my projects this week (pulled from session logs), what happened in my domains this week (pulled from external sources), and open questions I've been chewing on (pulled from my notes).

The Skill for this role is about 80 lines and includes the source list, the format spec, and a rule about staying under 1,500 words. I run this on Sunday at 5 PM. It takes the Researcher about 15 minutes. I review the output for 10 minutes and remove anything that's stale or irrelevant.

The output of this stage is raw material. Nothing publishable yet. But the entire newsroom downstream operates against this one document for the week.

Role 2: Angle Finder (30 min per cycle, weekly)

The Angle Finder reads the weekly research document and proposes 8-12 angles for the week. An angle is a one-sentence article concept, not a draft. Examples from a real week:

• "Why solo devs should run their own analytics layer instead of using Firebase"
• "I spent $1,200 on Apple Search Ads and got zero conversions — here's the autopsy"
• "Pricing tier surprises: why the cheapest tier lost"
• "How the Claude orchestrator pattern shipped my book in 8 weeks"

The Angle Finder is briefed to produce a mix: half technical, half opinion, half evergreen, half topical. It tags each angle with its target platform (Medium, LinkedIn, YouTube, Twitter) and its estimated audience fit.

I then pick three to four angles to actually pursue for the week. The Angle Finder produced too many on purpose — surplus options is a feature, not a bug. The 8-12 list gives me a real choice instead of committing to the first idea that comes up.

This is the only stage where I make creative calls. The rest of the newsroom executes on the choices I made here.

Role 3: Drafter (45-90 min per article)

The Drafter is one Claude session per article. For each angle I committed to, I open a new session, load the medium-draft Skill (or the linkedin-post Skill, or the youtube-script Skill depending on platform), pass it the weekly research document and the angle one-liner, and brief it on the take.

The brief is short. Three to five sentences. The angle, the why-this-matters-now, the data I want it to anchor on, and any constraint specific to the platform. The Drafter does the heavy lifting from there.

A Medium draft takes the Drafter about 15-20 minutes. A LinkedIn post takes 5-10 minutes. A YouTube script takes 30-45 minutes (longer because the YouTube format has specific pacing requirements that need more iteration).

I read every Drafter output. I correct course about 30% of the time after the first draft — usually with a "the opening's weak, try again from a stronger hook" or "this section is too theoretical, ground it in the specific QuranWay anecdote." Two or three iterations on average per piece.

The Skill for each platform contains the voice rubric, the structure constraints, and the negative examples I've collected over months. The Skills compound — each Friday review adds a new note to the relevant Skill if I notice the Drafter making a mistake systematically.

Role 4: Editor (15-30 min per article)

The Editor takes a Drafter output and runs three passes: a copy-edit pass (typos, awkward phrasings, inconsistent tense), a tone pass (does it sound like me?), and a structural pass (is the hook actually a hook? Is the closing actually a closing?).

The Editor's Skill is loaded with examples of good edits from my past work. When the Editor flags a passage as "voice drift," I can usually see exactly which earlier piece taught the Skill to recognize the pattern.

This role catches roughly 80% of the issues that would have made me uncomfortable shipping a piece. The remaining 20% I catch in my own final read. The Editor is not a replacement for human review; it's a replacement for the first round of human review.

Editor pass: 15-30 minutes per piece. The bottleneck is mine — I have to read the proposed edits and accept or reject. Most edits accept cleanly.

Role 5: Publisher (5-15 min per article)

The Publisher's job is to take the edited piece and prepare it for the destination platform. For Medium, it formats the markdown with the right frontmatter, picks the right header image from my asset library, sets the tags, and schedules the post via my Medium publisher Python script. For LinkedIn, it crops images to LinkedIn's aspect ratios, schedules via my LinkedIn poster. For YouTube, it prepares the description, the tags, and the thumbnail brief.

The Publisher is the only role that touches external APIs. It does not write content. It does logistics.

When the Publisher errors (cookies expired, image upload failed, schedule slot conflict), it surfaces the error immediately so I can fix it before the next piece in the queue. The Publisher Skill is the most boring of the five and the most reliable — it just runs the playbook.

The weekly rhythm

The newsroom runs on a weekly rhythm. Five touchpoints from me, total.

Sunday 5 PM (30 min): Run the Researcher. Review the weekly research document.

Sunday 6 PM (45 min): Run the Angle Finder. Pick three to four angles. Schedule them across the week's publishing slots.

Monday-Thursday morning (15 min each, 1 hour total): For each scheduled piece, do the brief to the Drafter. Walk away. The Drafter produces while I'm doing other work.

Monday-Thursday afternoon (30 min each, 2 hours total): Read the Drafter output. Decide on iterations. Hand to Editor. Quick review. Hand to Publisher.

Friday 4 PM (30 min): Weekly review. What landed well, what didn't. Update the Skills if any agent made a systematic mistake.

Total time: roughly four hours a week, distributed across the week. Output: four published pieces a week. About 16 pieces a month.

What this changes vs the alternative

The naive alternative is to write four pieces a week yourself. At my pace, that would take 12-15 hours a week of focused writing, which is genuinely not sustainable alongside running four iOS apps and a SaaS. Most solo creators try this for a month and then quietly cut back to one piece a week.

The five-agent newsroom moves the bottleneck from "write each piece yourself" to "make creative calls and approve outputs." The creative calls and approvals scale much better with time than the writing does. I can make five good creative calls in a Sunday evening. I cannot write five good drafts in a Sunday evening.

The compound effect over a year is enormous. 16 pieces a month × 12 months = roughly 192 published pieces. That is publisher-scale output from a solo operator. The audience numbers, the SEO compound, the reputation effect — all of those depend on the consistency of output, not on the brilliance of any single piece.

Where the chapter goes from here

The book chapter continues into the audience layer — how to grow distribution for the newsroom output (cross-posting, newsletter sequencing, audience segmentation), and the monetization layer — how the newsroom output ties back into the product business (lead generation for the SaaS, App Store referrals for the iOS apps, book sales).

Part 4 of the book covers three other one-person businesses built on the same agent pattern: the one-person dev shop, the one-person sales machine, and the one-person research lab. Each one uses the same chart-first, role-and-handoff approach. The newsroom is just the most visible example.

The book — How to Claude — The Playbook for a One-Person Business — has the full Part 4 walkthrough plus appendix C, which has the actual Skill files I use for the five newsroom agents, ready to copy into your own setup.

Or just start with one agent. Pick the role you find hardest to do consistently. Build the Skill. Run it weekly for a month. The newsroom builds from there.

Andy Garcia — solo dev, one-person newsroom, four App Store apps, one LegalTech SaaS. The full operating system is in the book.