One chat. One agent. One long context that fills with tool dumps, half-ideas, and "sure, let's build that." You call it brainstorming. The context window says otherwise... The problem When design and doing share one seat: - Context floods. Digests never form. - Weak approaches never die. They just get rewritten. - You "approve" a vibe, not a design. - Build starts before the frame can defend itself. Pretty threads. Soft decisions. Expensive rework. The idea: a design swarm I shipped a skill pack for this: brainstorming-swarm v1.0.0. Not a build swarm. A design swarm. 1. Fan independent lenses (failure, stakeholder, constraint, temporal, experiential). 2. Return digests only. Schema caps. No raw dumps into the conductor seat. 3. Hold hard gates for the human (redirect discovery, pick the winner). 4. Stop at a written design. Do not build from this skill. If your harness can run multi-seat, multi-tab, or multi-chat: dispatch-first. If it cannot: serial fallback with the same stages and honest labels. Floor is not default. The habit: orchestrator, not doer The main chat advises, routes, and synthesizes. Heavy work leaves the seat. Brief in. Digest out. Same turn when the work qualifies. Workers execute. They do not own taste or rewrite the method. You keep the gates. That is the job. Micro example Goal: redesign a personal reading list app. - Five lens digests, each capped. - Three competing approaches under different axioms. - Critique kills two, salvages one hybrid. - You pick. Spec writes. Session halts. No code in that run. That is the point. Get it Deep Dive (method + diagrams): https://aris-space.com/documents/workflows/brainstorming-swarm Members on Ari's Space: download the skill pack zip (v1.0.0). Load SKILL.md. Point it at a hard problem. Prefer dispatch-first. Stop at written design. Argue first. Code second. //A<3