Why we built 5 agents instead of one big prompt

The most common question we get is: "why not just send everything to Claude in one shot?"

Early versions of Inkwell did exactly that. One massive system prompt, one call, one document. The output was... fine.

But fine isn't the bar when someone's publishing a white paper or case study under their name.

The problem with a single prompt: it optimizes for everything at once and excels at nothing. SEO and readability are

in tension. Research depth and concise writing are in tension. You end up with averaged-out output.

So we split the pipeline into 5 specialized agents — Research, Writing, Formatting, Review, SEO — each with its own

system prompt, its own output contract, and no knowledge of what the others are doing. The Research agent doesn't know

there's an SEO agent downstream. The SEO agent only sees finished prose, not the research notes.

The result was noticeably better documents. But the bigger surprise was how much easier it became to iterate — if SEO

scores were low, we tuned one agent's prompt. If structure was off, we touched Formatting only. Single-responsibility

for AI agents turns out to work for the same reason it works in software.

Curious if anyone else has gone multi-agent vs single-prompt and what tradeoffs you hit. Happy to go deep on any part

of the architecture.