create-agent

name

create-agent

description

Use this skill when the user wants to scaffold a new Claude-based Python agent — single-agent loops or multi-agent coordinator systems built on the Anthropic SDK. Triggers include "create an agent", "scaffold an agent", "build me an agent that…", "I need a multi-agent system for…". Generates a runnable file that follows the course conventions (haiku-4-5 pin, four-category error schema, end_turn exit, optional coordinator + hooks). Do NOT use for editing existing agents — only for greenfield creation.

create-agent

Generates a new Claude-based Python agent that conforms to the Claude Certified Architect course conventions established in agent.py (single-agent loop) and capstone_project.py (multi-agent coordinator with hooks).

When to invoke

Trigger when the user asks to:

• "Create / scaffold / build an agent"
• "Make me a multi-agent system for X"
• "Set up a coordinator with subagents"

Do NOT invoke for: editing an existing agent, debugging tool calls, or explaining concepts. Those use direct edits / explanations.

How to run this skill

Follow these steps in order. Do not skip step 1 — the questions determine which template to use.

Step 1 — Gather requirements

Ask the user for these details. If they've already supplied any in their initial request, do not re-ask — only fill in the gaps.

1. Purpose — what is this agent for? (one sentence)
2. Architecture — single-agent or multi-agent (coordinator + subagents)?
3. Tools — list each tool's name, what it does, and its input parameters. For each tool, ask whether it's mock data (for teaching) or a real integration.
4. Hooks (multi-agent only) — does the user want any PreToolUse policy gates (e.g. spend caps, auth checks) or PostToolUse normalizers (e.g. raw → human-readable)?
5. Output filename — default to <purpose_snake_case>.py in the project root if the user doesn't specify.

Use AskUserQuestion when there are 2+ clarifying questions to bundle.

Step 2 — Pick the template

Architecture	Template
Single-agent	templates/single_agent.py
Multi-agent (coordinator + ≥1 subagent)	templates/multi_agent.py

Read the chosen template, then read reference.md for the non-negotiable conventions before writing.

Step 3 — Generate the file

Write the new file using the template as a starting point. Customize:

• Tool definitions in the tools / TOOLS array (name, description, input_schema)
• Tool implementations in execute_tool / TOOL_REGISTRY
• Mock data dicts (use the FAKE_DB / FAKE_ORDERS style from capstone_project.py)
• For multi-agent: add the user's hooks to PRE_TOOL_HOOKS / POST_TOOL_HOOKS, and edit the coordinator() function to decompose the task into the user's subtasks
• The if __name__ == "__main__": block to call the entry point with a sensible example

Step 4 — Verify

After writing, run a quick syntax check:

python -m py_compile <new_file>.py

If compilation passes, report the file path and the example invocation. Do not run the agent (it would hit the API and cost money) unless the user explicitly asks.

Non-negotiable conventions

These are enforced — do not deviate without explicit user approval. See reference.md for the full list with rationale.

1. Model: claude-haiku-4-5 only.
2. Error schema: tool errors must use the four categories transient | permission | validation | internal exactly as in agent.py.
3. Loop exit: stop_reason == "end_turn" is the only valid primary exit. MAX_ITERATIONS is a safety valve, not a stop condition.
4. Message order: assistant message with tool_use blocks is appended BEFORE the tool_result user message. Never reverse this.
5. No real secrets: mock data only. Read ANTHROPIC_API_KEY from env via anthropic.Anthropic() — never hardcode.

Files in this skill

• SKILL.md — this file (entry point)
• reference.md — the four conventions in detail, plus the message-shape rules
• templates/single_agent.py — minimal single-agent loop scaffold
• templates/multi_agent.py — coordinator + subagent + hooks scaffold