AI Agent Starter Kit

Everything you need to ship your first agent this weekend. Copy this structure, fill the blanks, and you have a working agent.

1. Project structure

my-agent/
  system-prompt.md     # who the agent is, goal, guardrails
  tools/               # one file per tool
  memory/              # notes the agent keeps
  examples/            # 5 real runs you graded

2. Base system prompt (fill the brackets)

You are [ROLE]. Your goal is to [GOAL]. You can use these tools: [TOOLS]. Always [MUST]. Never [NEVER] — if unsure, stop and ask. Before any irreversible action, show me a draft and wait for approval.

3. Your first three tools

• read — let the agent see one source of truth (a folder, a doc, a table).
• draft — let it produce output without sending anything.
• send — gated behind your approval at first.

4. Memory scaffold

Keep a short memory/facts.md of durable facts ("user prefers concise replies"). Prepend the relevant ones each run. That's 90% of memory you'll ever need.

5. Ship it

1. Run the agent on five real inputs.
2. Read every step — fix the prompt, not the model, when it stumbles.
3. If 4/5 are good, you have a useful agent. Use it Monday.

*Pair this with the Agent README Template and the Agent Evaluation Checklist.*