Multi-Agent Workflow Creator
A single AI doing everything produces generalist output at every stage. Multi-agent architectures assign specialized roles with explicit constraints — one agent generates, another critiques, another formats — producing higher quality at each handoff point than any single prompt can achieve.
The Recipe
Act as a Principal Systems Architect and AI Automation Engineer. I want to build a highly efficient, automated multi-agent workflow to execute a complex, recurring project without human bottlenecking.
The overarching project/workflow I want to automate is: [INSERT DETAILS, e.g., A system that takes a raw codebase file, security audits it, documents it, and writes unit tests].
Please design a complete "Multi-Agent System Architecture Blueprint" including:
- Agent Roster & Personas: Define 3 distinct AI agents required for this pipeline. Give each a highly specific role, an explicit skillset, and a psychological constraint (e.g., Agent B is a ruthless cynic who only looks for flaws).
- The Hand-Off Logic Model: Map out the sequential and conditional data flow. Explain exactly what trigger or file state causes Agent A to pass its output to Agent B.
- System Context Prompts: Provide the exact, optimized system instructions for the single most critical agent in this sequence to ensure zero hallucination and maximum technical execution.
The three blueprint components
| Component | What it produces |
|---|---|
| Agent Roster & Personas | 3 agents with specific roles, skillsets, and behavioral constraints — not generic helpers |
| Hand-Off Logic Model | The conditional triggers that move output from one agent to the next, including failure states |
| System Context Prompts | The full system prompt for the most critical agent — ready to deploy |
Psychological constraints — why they matter
A “ruthless cynic who only looks for flaws” produces different output than a “helpful assistant who also checks for issues.” Psychological constraints force agents into specific cognitive modes:
- Critic persona — surfaces problems a neutral reviewer misses
- Paranoid security mindset — finds edge cases a standard reviewer accepts
- Rigorous skeptic — questions every assumption rather than building on them
The constraint prevents the agent from drifting toward agreeable, generic output.
Hand-Off Logic — the most underspecified part of most workflows
“Agent A passes to Agent B” is not a hand-off model. A proper hand-off specifies:
- What exact output format triggers the handoff (a file, a JSON object, a status flag)
- What condition causes Agent B to reject the handoff and send it back
- What happens at failure — does the pipeline stop, loop, or escalate to human review?
🔁 Leftover Remixes
🌶️ Spicy: “After the blueprint, identify the single highest-failure-risk point in this pipeline — the step most likely to produce hallucinations or bad output — and give me a validation prompt to catch errors before they propagate downstream.”
🧊 Mild: “Skip the full blueprint. Just give me the Agent Roster and psychological constraints — I’ll design the hand-off logic myself.”