Guides

The 7 steps

1. Pick one boring job. Choose a task you do at least weekly that follows the same pattern each time: sorting email, writing product listings, summarizing patch notes, quoting a print job. One job per agent. Agents that try to do everything do nothing well.
2. Write the rules in plain English. List what good output looks like, what the agent must never do, and what it should do when it is unsure. Example: 'Sort into Urgent, Reply, Read later, Archive. Never send anything. If you can't tell, put it in Reply.'
3. Give it two or three real examples. Paste a couple of real inputs with the output you would have written yourself. Examples teach format and tone faster than any instruction.
4. Put it somewhere it remembers. Save the rules as a ChatGPT Project or custom GPT, a Claude Project, a Grok workspace, or a rules file in Cursor. That way you don't re-paste the setup every time.
5. Keep a human in the loop. Have the agent draft, sort, or suggest, and you approve. Don't let a new agent send messages, spend money, delete things, or post publicly on its own.
6. Run it for a week, then tighten. Every time it gets something wrong, add one rule or one example that would have prevented it. Most agents get good after five to ten small fixes.
7. Automate only after it is reliable. Once the output is consistently right, connect it to a trigger with a tool like n8n, Zapier, or your agent app's schedules, so it runs without you starting it.

A starter template you can copy

You are my [job name] assistant. Your only job is [one sentence].

Rules:
- Output format: [exact format]
- Always: [2-3 must-dos]
- Never: [send, delete, buy, post, or anything else off-limits]
- If you're unsure: [ask me / flag it / put it in a 'check' pile]

Examples:
Input: [real example 1]
Output: [what you would have written]

Input: [real example 2]
Output: [what you would have written]

Common first-agent mistakes

• Asking one agent to do five different jobs.
• Letting it send or post before you've checked a week of drafts.
• Vague rules like 'be helpful' instead of concrete ones like 'max 3 sentences'.
• No examples, so the format changes every run.
• Automating before the manual version works.

Ideas for your first agent

Browse the free AI agent use case directory for jobs with setup steps and starter prompts, or jump to a niche: productivity, 3D printing, gaming, vibe coding. Good first picks: email triage, Etsy/eBay listing writer, and error explainer for vibe coders.

Want it done for you?

Each Bot Pantry kit is this process already finished for one job: tested rules, examples, setup guides for popular tools, and a safety checklist.