ai-agent-designer

name	ai-agent-designer
description	Designs AI agent architecture for Make.com scenarios. Determines model selection, tool inventory, memory strategy, looping pattern, and output contract. Called by scenario-orchestrator during System Design phase when ai_required is true, or directly by ai-agent-builder agent.

Skill: ai-agent-designer

Designs AI agent architecture inside Make.com before any blueprint is written. Called when an automation requires LLM reasoning, classification, generation, or decision-making.

Sub-files

• AGENT-DESIGN-INTAKE.md — Interview protocol (5 questions), model selection matrix, tool inventory
• AGENT-DESIGN-BLUEPRINT.md — Blueprint output format, prompt engineering rules, output contract

Deterministic Classification

This skill makes ONLY read-only MCP calls during design. No scenarios_create, no scenarios_update — those happen in Sprint phase only.

Permitted: AskUserQuestion, apps_recommend, app-modules_list, app-module_get, app_documentation_get, data-stores_list, Read / Write

Precondition — Routing Verdict

Do not run this skill without a Routing Verdict from skills/execution-model/SKILL.md.

• ai-agent or hybrid → proceed. On hybrid, design the agent for only the indeterministic step named in the verdict, and give it a structured output contract so everything downstream is deterministic again.
• scenario → stop. Tell the user the automation is deterministic, cite the verdict's Reason line, and hand back to plan-builder. If they have information that changes the picture, re-run execution-model with it — never override the verdict silently.

For current Make AI Agent semantics — tool calling, memory, iteration limits — load skills/help-docs/SKILL.md rather than answering from memory; this surface changes often.

When to Call This Skill

The keyword list below is a detection hint for the interview, not the decision. The decision is the Routing Verdict above.

Call this skill whenever an automation contains any of:

• "AI", "LLM", "language model", "ChatGPT", "Claude", "Gemini"
• "summarize", "classify", "generate", "decide", "extract", "respond intelligently"
• "AI agent", "chatbot", "autonomous", "tool use"
• "analyze this email and decide", "write a reply", "determine whether..."

If detected during kickstart-intake: mark ai_required: true in the automation object. During Phase 2, the orchestrator calls this skill before plan-builder.

Integration with Other Skills

After this skill completes:

1. Call ai-docs-researcher to look up exact Make.com module specs for the chosen model
2. Call agent-pattern-library to load the full pattern blueprint
3. Hand AI Agent Blueprint to plan-builder as additional context
4. The plan-builder incorporates the AI steps into the full AutomationPlan