Build your first AI agent.

You ask a chat, and it answers. Tomorrow you start over. You set an agent up once. After that, you hand it the week's notes and it hands back a draft.

The tools change every month. The building blocks stay put. This is the handout from the Flathead AI Lab, Mark Ulett's hands-on workshop: a small AI agent in Claude or ChatGPT that does one job you repeat every week. The free plans are enough for memory and instructions, which is where you start. In Claude a schedule needs a paid plan. In ChatGPT the free plan allows three scheduled tasks, each once a day at most.

Do this on a computer, in a web browser such as Chrome or Safari. The phone apps name things differently. In both tools, Settings is in the menu that opens when you click your name or picture, usually at the bottom left.

Start here: pick one job

A job you do every week, small enough to start today. You give it something you paste in. It hands back something you check and send.

Paste this into Claude or ChatGPT:

Help me pick one job I do every week that an AI agent could draft for me. Ask me three questions, one at a time. Then write it as one sentence: When I give it ___, it hands me back ___.

My agent's job: When I give it ___, it hands me back ___.

Make a project for it. A project is a folder that keeps one job's chats, files and instructions together. In Claude, click Projects in the sidebar (the list down the left side of the screen), then + New Project. In ChatGPT, click New project in the sidebar. Give it the name of the job. Open a new chat inside the project. The project is your agent, and you talk to it in that chat.

Six blocks make an agent

1. Memory. What it knows about you.
2. Instructions. Who it is and how it works.
3. Skills. How to do each task.
4. Schedule. When it runs on its own.
5. Permissions. What it may do without asking.
6. Connectors. The apps it can reach.

Start with memory and instructions. Add the rest as you need them.

1. Memory

What it is: what the agent learns about you and keeps between chats.

• Claude: open Settings, then Your Claude (older versions say Memory), and turn on Generate memory from chats. Each project keeps its own memory.
• ChatGPT: open Settings, then Personalization, then Memory, and turn it on. To keep a project separate from your other chats, set its memory to Project-only in the project settings.

Help pages: Claude on memory and ChatGPT on memory.

Try it. Tell your agent:

Remember that I want drafts short, in plain words, with the key point first.

Later, ask it:

What do you remember about me in this project?

2. Instructions

What it is: who the agent is, its one job, and how it treats you. You paste this once on the project page, and every chat starts from it.

In Claude, click Set project instructions on the project page. In ChatGPT, open the three dots menu (•••) in the project, then Project settings.

Help pages: Claude on projects and ChatGPT on projects.

Copy this, fill in the blanks, and paste it into your project instructions:

You are my [job] helper.

Your one job: when I give you [what I paste in], you hand me back [what I check and send].

How you work:

If something is missing, ask me one question. Don't guess.

Keep it short and plain. Put the most important thing first.

You draft. I check it and I send it.

Where to look: the files in this project and the examples I give you.

How you treat me: [paste the soft doctrine below]

The soft doctrine

The soft doctrine is ten house rules for how your agent treats you. It goes into your project instructions, in three steps:

1. Paste the instructions box above into your project instructions.
2. Paste one tone line from the next box over the words [paste the soft doctrine below].
3. Paste the ten rules under it.

Then change any line you like.

Pick one tone and delete the other two:

Gentle: Be warm and reassuring. Explain things in plain words, and point out what I got right.
Straight: Be friendly and get to the point. Skip the cushioning.
Sassy: Be blunt and a little playful. Push back hard when I'm wrong. Never be mean, and never get sloppy with facts.

How you treat me:

1. Start from what I need. Work out what I'm really trying to get done. If you can't tell, ask me one short question.
2. Use what I told you first. My facts about my work and my voice come before anything general. When you fill in with general knowledge, say so.
3. Say what you don't know. Never fill a gap with a guess that looks like a fact. Never make up names, numbers, dates, prices or quotes.
4. I decide, and I send. You draft, suggest and flag. Before anything that can't be undone, stop and ask me.
5. Tell me the hard thing kindly. If my plan has a hole, say so once and clearly. Then help with whatever I choose.
6. Be honest about your work. Only say you did something if you did it. If a step failed, tell me.
7. Keep us both on track. If we drift from the job, say so and offer to get back to it.
8. Look after me, not just the task. If I seem rushed or stuck, make the next step small. For legal, money, health or safety questions that matter, point me to a real person.
9. When I correct you, keep it. Use my correction for the rest of the chat, and remind me to add it to these instructions.
10. Give the same care every time. Treat my hundredth request like my first.

Test it. Give your agent one real example and one tricky one, with something missing. Save both in a note. Run them again when the tool changes: the companies swap in a new version every few months, and answers can shift.

3. Skills

What it is: how to do one task, the same way every time. One skill for each task, so you can fix one without touching the rest. A job can have two or three tasks in it.

• Claude: skills work on every plan, the free one too. Type "Turn this into a skill". Under it, paste your own version of the example in the box below. Click Save skill on the file Claude makes. It then shows under Customize (in the sidebar), then Skills. If Claude cannot make the file, open Settings, then Capabilities, and turn on Code execution and file creation (newer versions call it Create and edit files).
• ChatGPT: skills are on business and school plans, not on personal ones. If you have them, open Plugins (it is in the sidebar, and also under Settings), then the Skills tab. On a personal plan, add the skill to the project as a file: the project page has a place to add files, and you can paste the text in.

Help pages: Claude on skills and ChatGPT on skills.

Every skill has the same five parts. An example to copy and change:

What it is for: turn my notes from the week into a short summary.

What I give it: my notes from the week, pasted in.

How to do it: pull out the decisions, the open questions and the next steps. Write five bullets or fewer, in plain words. End with one question for me if anything is unclear.

What good looks like: "Decided: we open Saturdays in December. Open: who covers the 14th? Next: Jen orders the new sign by Friday."

Never: never add anything that isn't in my notes. If something is unclear, ask me.

4. Schedule, if you want one

What it is: when the agent runs on its own. Add it last, once you trust the agent, and read what it sends you the first few times.

• Claude: click Scheduled in the sidebar. It needs a paid plan.
• ChatGPT: click Scheduled in the sidebar. The free plan allows three tasks, each once a day at most. A task can't read your project files, so paste your instructions into the task itself.

Help pages: Claude on scheduled tasks, in its Cowork guide and ChatGPT on scheduled tasks.

An example task to copy and change:

Every Monday at 7:00 AM: find three news stories from the past week about [my topic]. For each one, give me the headline, two plain sentences on why it matters to me, and the link. Keep the whole thing under 200 words.

5. Permissions

What it is: what the agent may do without asking you first.

Nothing to set until you connect an app in block 6.

• Claude: each connector has its own settings: Always allow, Needs approval or Blocked. Find them under Customize, then Connectors, then the connector.
• ChatGPT: open Settings, then Plugins (older versions say Apps), then the app's permissions. Choose Allow read actions. Then it can read on its own, and it asks before it changes anything.

Help pages: ChatGPT on app permissions.

Rule of thumb: connect only what is safe to connect (block 6 says how to tell). Then let it read, and make it ask before it writes, sends or deletes. It drafts. You send.

6. Connectors

What it is: the apps your agent can reach, such as mail, calendar and files.

A connected app lets it read what is in there, customer details too. Read What should I never paste into an AI tool? before you connect one. If your mail or files hold what is on that list, do not connect them.

• Claude: open Customize, then Connectors, and pick one: Gmail, Google Calendar, Google Drive and more.
• ChatGPT: it calls these apps. Open Settings, then Plugins (older versions say Apps). Pick the app, click Connect, and sign in.

Help pages: Claude on connectors and ChatGPT on connecting apps.

Tip: connect one app. Add more once you trust it.

Before you go

• Your agent is your files. Keep a copy of your instructions and skills where you can find them.
• Check on it when the tool changes. The companies swap in a new version every few months. Run your two test examples again and compare.
• Let it go when its job is done. Agents come and go. The building blocks stay.

It drafts. You send.

When the menu names were checked

The menu names here are from the Lab of November 3, 2026, and were checked against Claude's and ChatGPT's own help pages on November 8, 2026. They change. If one looks different, the help page beside that step has the current name.

What does a finished AI agent look like? has a complete example, laid out in all six blocks.