Sub-Agents vs One Big Prompt: Why Division of Labor Beats a Genius Assistant

New users write one giant prompt: "You are a marketing expert that does research, writes content, checks facts, optimizes for SEO, and formats for Instagram." The result is mediocre at everything.

The reason is simple: large-language models, like humans, lose focus when juggling too many roles in a single context window. Quality drops on the last 3 tasks while the AI satisfies the first 2.

Split the work. Each sub-agent is a separate Markdown file describing one role: research_agent.md, content_agent.md, review_agent.md. Each has its own tools, its own output format, its own success criteria.

The main CLAUDE.md orchestrates: "If the user asks for a weekly newsletter, first call research_agent, then hand its output to content_agent, then hand that to review_agent for fact-check and brand compliance."

Each agent runs with a small, focused context — and produces better output than the mega-prompt.

You don't need five agents. Start with three: a researcher (gathers sources), a writer (drafts output), a reviewer (fact-checks and brand-checks).

The Harness Builder wizard creates these three files automatically from your answers. Each file is about 30 lines.

Upfront, yes — about 10 extra minutes on the first setup. Ongoing, it's far less work: when output quality drops, you tune one agent file instead of rewriting a mega-prompt and losing everything that worked.