Tools of the trade

A chat assistant only sees what you paste in. An agent works in your files, runs commands, and acts with your keys.

• It acts as you.It signs in with your account, so it can do what you can do. Connectors like Gmail, Drive and Calendar ask for access on a consent screen. Read that screen: it lists every key you hand over.
• Read → draft → send → delete.Each step up is a bigger key. Start at read. Add a step only after you've watched the one below it work.

The email ladder

It asks before it acts

Both products have something called “Auto,” and they mean different things. Start where it asks the most, and loosen one notch at a time.

• What leaves your laptop.Whatever the agent reads goes to the model. A cloud run happens on the company's machines, with whatever you connected to it.
• Write the house rules down.A plain-text SOP the agent reads first. See the real one below.

This is the file in today's demo folder. No code in it: an SOP in plain English.

# House rules for this folder

Everything in this folder is dummy data made for a workshop. Treat it as if it were real.

- Work only inside this folder. Don't read or write anywhere else.
- The CSV files you were given are originals. Never edit them; write new files next to them.
- Before you run any code, explain in plain English what it will do.
- Put exact checks (duplicates, blanks, formats, dates) in code. Use your own judgment only where I ask
  for it, and say which is which.
- When a row doesn't fit the rules, mark it "needs review" with a reason. Don't guess.
- Never send email, post, or upload anything.

Five ways, from “you run it” to “it runs itself.” The further down, the less you watch.

1. A

   By hand, in a session

   You ask, you watch every step, you say yes or no. The smallest useful unit starts here.

   Claude Code's /loop repeats a prompt inside an open session until a goal is met, for up to about three days.

   Laptop on, you're at itClaude Code · Cowork · Codex
2. B

   A trigger inside a tool you already use

   Google Apps Script. On form submit runs when someone responds to a form. Time-driven runs anywhere from every minute to once a month; a 9 AM trigger fires between 9 and 10, then keeps that time.

   It runs as the person who created the trigger, so any email it sends comes from that account.

   Laptop can be closedToday's hands-on, Track A
3. C

   Scheduled on your own computer

   cron runs a command on a timetable. On a Mac, Apple recommends launchd instead. The command can be a script, or an agent run without a chat window: claude -p "…" or codex exec "…".

   Only while the computer is on and awake. A launchd job missed during sleep runs when the Mac wakes; a cron job waits for its next time.

   Laptop must be on
4. D

   Scheduled tasks built into the apps

   • Claude Code, desktop scheduled tasks: on your machine, awake with the Desktop app open. As often as every minute, with your local files and a permission mode for each task.
   • Claude Code routines: in Anthropic's cloud. On a schedule (at most hourly), from an API call, or on GitHub events. No local files. Set up at claude.ai/code/routines or with /schedule.
   • Claude Cowork scheduled tasks: /schedule. They run in the cloud and keep going when the Desktop app is closed, but reach your connected folders only while it is open.
   • Codex scheduled tasks (formerly Automations): under Scheduled in the ChatGPT desktop app, which works as an inbox. On the desktop they can use your project folder, with the computer on and the app running. Made on the web, they use uploads and connected tools, not your folders.

   Laptop: depends where it runs
5. E

   In the cloud: GitHub Actions

   A workflow file in a GitHub repository, run on GitHub's machines on a schedule or on events such as a pull request. Claude Code and Codex each have an official Action: anthropics/claude-code-action and openai/codex-action.

   API keys live in repository secrets. Every run uses the key, and API use costs money.

   Laptop can be closed

Side by side

# minute hour day-of-month month day-of-week  command
0 9 * * 1  /Users/you/intake/run-triage.sh

name: Weekly check
on:
  schedule:
    # By default, cron time is UTC. 13:00 UTC is 9:00
    # in Michigan during daylight saving time.
    - cron: '0 13 * * 1'
jobs:
  check:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - run: python3 check.py
        env:
          API_KEY: ${{ secrets.API_KEY }}  # a repository secret

7 of 11 client projects involve sensitive data or an NDA. Answer four questions.

1. What do I need to connect?
2. What tools have access to those systems?
3. What is the simplest approach that works?
4. Can an AI tool help me bridge the gap between systems?

Three scenarios

What is an API?

Application Programming Interface: a list of things a system can do, and the rules for how to ask it to do them.

How do computers talk?

JSON: a list of keys paired with values.

{
  "name": "John",
  "role": "Manager",
  "department": "Operations"
}

Sound familiar? That's the email ladder.

Where AI comes in

LLMs translate human language into computer language.