Multi-Agent Content Factory

Solo content creators spend most of their time on inputs (research, ideation) and outputs (formatting, scheduling) — not actual writing. Multi-agent delegation parcels out the boring work to specialised subagents and lets you focus on the parts that need your voice.

What it does

Researcher subagent — given a topic, returns a structured brief (key facts, sources, angle suggestions)
Writer subagent — given a brief, drafts a first version in your voice
Editor subagent — critiques the draft against your style guide, surfaces weak claims and clunky phrasing
Publisher — formats final approved version for the target platform (blog, newsletter, X thread)
All four coordinated via the orchestrator (the agent you talk to directly)

What you'll need

Delegate tool — for spawning subagents
A style-guide skill capturing your voice (tone, sentence length, banned words)
Memory for the topic + draft state
Access to whichever publishing channel (RSS, ghost API, email service, X)

Setup

1. Write your style guide

Create ~/.flowly/skills/style-guide/SKILL.md:

markdown

---
name: style-guide
description: Use when writing or editing any content I publish.
---

# Style guide for Hakan's writing

## Voice
- First person, conversational but specific
- Sentences average 18 words. Vary length.
- No hedging filler: avoid "in many ways", "it could be argued"
- Concrete > abstract: name a thing instead of describing it

## Banned phrases
- "delve into", "tapestry", "in the realm of"
- "robust", "leverage" (as a verb)
- Any LLM-favourite intensifier: "incredibly", "remarkably"

## Structure
- Opening line is a hook, not a setup
- Tight paragraphs: 2-4 sentences
- End sections with a forward-pointing line ("next, …")

2. Define the workflow

Send to Flowly

When I say "draft a post about X": 1. Spawn researcher subagent: "Research X for a 1500-word piece. Return: - 5 specific facts with sources - 2 contrarian angles - 3 anecdotes or examples I could use" Wait for completion. Save to memory tagged "draft:<slug>:research". 2. Spawn writer subagent: "Write a 1500-word draft using the style-guide skill and the research above. First-person, hook in line 1. Title is a bonus." Wait. Save to "draft:<slug>:v1". 3. Spawn editor subagent: "Critique v1 against style-guide. List: - 3 weakest sentences and why - Any banned phrases or hedging - Whether the hook actually hooks - Specific revision suggestions" Wait. Save to "draft:<slug>:critique". 4. Send me the v1 + critique. I'll either ask for v2 (with my feedback) or approve. 5. On approval, spawn publisher subagent: "Format draft v1 for [target]. For blog: HTML with semantic headings, alt-text on images. For X thread: 280-char tweets, threading marks. For newsletter: subject line + plaintext body." Each subagent runs in parallel where possible (researcher and style-guide can load simultaneously).

3. Try it

"Draft a post about why most AI agent demos fail in production."

The agent kicks off the pipeline. Within 5–10 minutes you have a draft + critique. Iterate from there.

Tips

Style guide is the highest-leverage piece. A tight, specific style guide gets you 80% of the value. Vague style guides ("write conversationally") produce vague drafts.
Don't skip the critique. First drafts from any LLM read like first drafts. The critique surfaces 3 things you'd have caught yourself but faster.
Researcher needs sources. Bake "with citations" into the researcher prompt. Without it, you get plausible-sounding inventions.
You're still the writer. The pipeline produces a starting point. Your voice goes in via heavy edits. Output that's 100% AI reads like 100% AI.
Watch token costs. A full pipeline run is 4 LLM calls × maybe 10k tokens each. For your blog cadence, that's fine; for high volume, switch researcher and editor to a cheaper model (Haiku, Kimi).