agent-design

name

agent-design

description

design and review llm agent systems, tool interfaces, and skill boundaries. use when chatgpt needs to decide whether a task should be handled by a prompt, workflow, skill, tool, single-agent loop, or multi-agent system; define tool schemas, instructions, guardrails, evaluation plans, or permission boundaries; or critique an existing agent design for over-orchestration, weak context handling, missing verification, or poor tool design.

Agent Design Skill

Core Principles

Three axioms everything else derives from:

1. Tools are primitive system boundaries. They expose raw capabilities without encoding usage assumptions.
2. Skills are usage knowledge. They teach the agent when and how to wield those primitives—as knowledge, workflows, CLI wrappers, delegation patterns, or tool integration guides.
3. The context window is real state. Everything else is a projection.

Quick Decision Rules

Tool or Skill?

• New system boundary (shell, filesystem, web, external API) → Tool
• Teaching the agent to use an existing capability reliably → Skill
• See decision-tree.md for full breakdown

Workflow or Agent loop?

• Subtasks are predictable and fixed-order → Workflow
• Path depends on prior results and can't be hardcoded → Agent

Single-agent or Multi-agent?

• One context window is sufficient → Single
• Tasks are genuinely independent, or require specialized context separation → Multi
• When in doubt, start single. See patterns.md

When to add orchestration?

• Only when simpler structures demonstrably fail. Never for architectural elegance alone.

Can I make a skill always on?

• No. Skills are not permanently active by default. They are activated by relevance.
• Do not design a skill as a hidden, always-on system prompt section. If you truly need global behavior, make the activation rule explicit, for example: Always load the xxx skill first.
• If you want an always-on skill, state that intent clearly in both the system prompt and the skill description, for example: Load this skill unconditionally, regardless of the user's current request.

Red Flags

If you see any of these, the design has likely drifted:

• A tool that encodes "how to use it" in its interface → Interface Wrapping
• An agent that reports its state more than it acts → Meta-Reporting
• A coordinator that every agent must ask before doing anything → Orchestration Gravity
• Approval gates on every meaningful action → Safe Mode Trap
• External database treated as the agent's real state → State Externalization Fallacy
• System prompt rebuilt every turn with current state → Append-System-Prompt

References

Topic	File
When to use Tool vs Skill vs Workflow vs Agent	decision-tree.md
Tool design & ACI	tool-design.md
Context, state & session management	context-and-session.md
Orchestration & structural patterns	patterns.md
Behavioral anti-patterns	anti-patterns.md
Evaluation & verification	evals-and-verification.md
What is Skills?	What-is-skills.md