Executive Ops AI

RESOURCES / GUIDES / AI BASICS

GUIDE BEGINNER ABOUT 45 MINUTES

Set up your first AI agent.

Learn what an AI agent is, choose a good first job, and build one in Claude, step by step. By the end, a draft of your weekly brief will be ready for you every Friday, without you having to ask.

Start the setup Watch the film first
YOUR “RECIPE”THE DRAFT 1234 THE JOB FRI 12:00 A PERSON APPROVES
YOU NEED
A paid Claude plan
SETUP TIME
About 45 minutes
LEVEL
Beginner, no coding
LAST CHECKED
26 September 2026
SPEED

WATCH FIRST · 2:49

What is an AI agent?

Kelly runs operations for a leadership team. The film follows one Monday meeting to show what an agent does, where the limits go, and where people stay in charge.

01 / 10 Chatbot or agent
  1. 01Chatbot or agent
  2. 02Words you’ll hear
  3. 03Choose a first job
  4. 04Pick your path
  5. 05Before you start
  6. 06Build it step by step
  7. 07Keep it safe
  8. 08What to expect
  9. 09Ideas to try
  10. 10When you’re ready

01 · THE IDEA

A chatbot answers. An agent does the steps.

Both run on the same kind of AI. The difference is what happens after you ask. A chatbot writes a reply and stops there. An agent can take actions, one after another, until the job is done.

A CHATBOT

You ask. It answers. You do the rest.

  1. You paste in what it needs.
  2. It writes a reply.
  3. You copy the reply somewhere and finish the job yourself.

Good for drafting, explaining and thinking out loud.

AN AGENT

You give it a goal. It works through the steps.

  1. It plans what to do first.
  2. It uses the apps you have connected.
  3. It checks its own work and hands you a draft to review.

Good for work with several steps that comes round every week.

PlanPICK THE NEXT STEPDoUSE A TOOLCheckDID IT WORK?AdjustFIX AND RETRY
An agent works in a loop. It plans a step, does it, checks whether it worked, and tries again if it did not, until the job is done or it needs your help. That loop is what lets it handle real work, and it is also why it needs clear limits.

02 · VOCABULARY

Ten words you’ll hear.

No need to memorise these. Skim them now, and come back if a word in the steps is new to you.

Agent
An AI that can take steps toward a goal, as well as answer questions.
Prompt
What you ask for, written out. Clear instructions get better results.
Connector
A secure link that lets an AI tool read from, or act in, another app, such as your calendar or shared files.
Permissions
What the agent may do in each app: read, change, or nothing at all. You decide.
Approval mode
How often the agent stops to ask you before it acts.
Scheduled task
A job that runs by itself on a timetable, such as every Friday at noon.
Skill
A saved set of instructions the agent picks up when a job needs it, so the work is done the same way every time.
Project
A workspace that keeps background files and standing instructions for one area of work.
Hallucination
When AI states something false and sounds sure of it. The reason a person checks the result.
Prompt injection
Hidden instructions inside a web page, email or file that try to make an agent do something you never asked for. Chapter 07 explains it.

03 · CHOOSE WELL

Start with a small job you already do every week.

Pick something you already do by hand, where you would catch a mistake before it caused a problem. Then ask yourself these three questions about it.

1

Does it come round every week?

Repeating work pays back the setup time. A one-off job is quicker to do yourself.

2

Could you write the steps on an index card?

If you cannot explain the steps, the agent cannot follow them.

3

Can you check the result in a few minutes?

You will read every result for the first few weeks. Pick something you can check quickly.

Three yeses make a good first job.

GOOD FIRST JOBS

  • Gathering information from a few places into one summary.
  • Drafting something you will review: a brief, an agenda, a recap.
  • Sorting and tidying: meeting notes into actions, receipts into a table.
  • Spotting gaps: meetings with no agenda, tasks with no owner.

04 · THREE WAYS IN

Three ways to start. Each one builds on the last.

You can stop at any path and still have something useful. The example in this guide goes through all three, so you can see how a simple task grows into one that runs on its own.

  1. PATH 1

    Try a task

    Open Claude and hand it one job, pasting in what it needs. Nothing is connected and nothing runs on its own.

    TIME
    10 minutes
    YOU NEED
    Claude on the web, desktop or phone

  2. PATH 2

    Connect one app

    Let Claude read your calendar or a shared folder so you stop copying and pasting. Start with read-only access.

    TIME
    20 minutes
    YOU NEED
    A paid plan and your company’s OK

  3. PATH 3

    Put it on a schedule

    Save the job as a scheduled task so the draft is ready for you every week without being asked.

    TIME
    15 minutes more
    YOU NEED
    A paid plan and a connected app

Other assistants, such as ChatGPT and Microsoft Copilot, have similar agent features. The ideas on this page carry over; the menus will be different.

05 · GET READY

Five things to have ready.

Tick them off as you go.

BEFORE YOU START

0 OF 5

06 · BUILD IT

Build a Monday brief agent, step by step.

The example follows Kelly from the film: every Friday, a draft one-page brief for Monday’s leadership meeting. Swap in your own job as you go. The steps stay the same.

  1. Step 1. Check you are allowed.

    5 MIN

    WHYSo work information only goes where your company has approved.

    Find your company’s AI policy, or ask IT which AI tools are approved and which account to use. If Claude is not approved, stop here and send this guide to whoever decides.

    YOU ARE READY WHEN

    • You know which account to use.
    • You know which apps you may connect.
    • You know which information must stay out of AI.

  2. Step 2. Install Claude and set it to ask first.

    5 MIN

    WHYWatching every action is the fastest way to learn what the agent does.

    Download the Claude desktop app from claude.com/download and sign in with the approved account.

    Beside the message box, set the approval mode to Manual. Claude will now stop and ask before every action. You can loosen this later, one job at a time.

  3. Step 3. Make a practice folder.

    5 MIN

    WHYA small folder keeps any mistake small.

    Create a folder called Agent practice and copy in a few files that are safe to share: last week’s meeting notes, a project list, a sample agenda. When Claude asks which folder it may use, choose this one and nothing else.

  4. Step 4. Write the job down as a “recipe.”

    10 MIN

    WHYClear instructions are the biggest difference between a useful agent and a frustrating one.

    By “recipe” we mean your written instructions for the job. You write them once, and the agent follows them the same way every time, the way a cook follows a recipe.

    A good “recipe” has five parts: the goal, where to look, the steps, what the result should look like, and the rules. The last rule matters most. Tell it to ask instead of guessing.

    YOUR “RECIPE”

    Goal: Draft a one-page brief for Monday’s leadership meeting.

    Use: the files in the Agent practice folder.

    Steps:

    1. List the three most important project updates, with the owner of each.

    2. List anything that is overdue or has no owner.

    3. Suggest a short agenda: numbers, projects, decisions needed.

    Format: one page with three headed sections, in plain sentences. After each fact, put the name of the file it came from in brackets.

    Rules: Do not send, delete or change anything. If something is missing or unclear, list it under “Questions for me” instead of guessing.

  5. Step 5. Run it once and watch.

    10 MIN

    WHYThe first run shows you how it thinks, and where your “recipe” is unclear.

    Paste the “recipe” into Claude and send it. Claude shows its plan, then asks before each action. Read what it wants to do before you approve it.

    When the draft arrives, check it against the files. If something is wrong, fix the “recipe” rather than the draft, so it gets it right next time.

    A GOOD FIRST DRAFT

    • Every fact has a file name in brackets.
    • Nothing was invented to fill a gap.
    • The questions section lists what it could not find.

  6. Step 6. Connect one app, read-only first.

    5 MIN

    WHYA connection saves the copying and pasting, and read-only access means it can look without touching.

    Open Customize, then Connectors, and connect the app that holds your material, such as Google Calendar or Google Drive. Where the app offers a choice, allow reading only, or set anything that changes, sends or deletes to need your approval.

    Then add one line to your “recipe”: “Also use: next week’s calendar, read only.”

  7. Step 7. Put it on a schedule.

    5 MIN

    WHYOnce it is scheduled, the draft turns up on its own. You no longer have to remember to ask.

    Choose Scheduled in the sidebar, then New task, then Create with Claude. Paste your “recipe”, choose every Friday at 12:00, and keep the approval mode on Manual for now. You can also type /schedule in the message box.

    IT USES CONNECTED APPSIT USES FILES ON YOUR LAPTOP zz Runs with the lid closed. Runs only when the laptop is on and Claude is open.

  8. Step 8. Review the first three runs.

    5 MIN A WEEK

    WHYA few good runs in a row is how you know it is doing the job properly.

    Read each draft before Monday and note what you had to fix. After three good runs, decide whether anything can be loosened, such as switching from Manual to Auto for a job that only reads.

    LOOSEN IT ONLY WHEN

    • It ran on time three weeks running.
    • The facts matched their sources every time.
    • You changed less each week.

WHAT ARRIVES ON FRIDAY

The draft, waiting for you.

  1. 1Every fact carries the file it came from, so checking takes minutes.
  2. 2It flagged the item with no owner instead of inventing one.
  3. 3The last section is a good sign. A well-set-up agent tells you what it could not find.

MONDAY BRIEF · DRAFT FOR KELLY

PREPARED FRIDAY 12:04

1PROJECTS

Office move: on track, owned by Dana. [project-list]

Q4 hiring plan: waiting on budget sign-off. No owner listed. [notes-sep-18]

2OVERDUE

Vendor contract renewal was due on 20 September. [project-list]

DECISIONS NEEDED

Hire now, or wait until May? [notes-sep-18]

3QUESTIONS FOR ME

The notes mention a board date but not the day. Which date should I use?

07 · SAFETY FIRST

Five rules that keep an agent safe.

An agent uses your accounts, so you are responsible for what it does. These five rules keep the risk low while you are learning.

  1. 1

    Start read-only.

    Let it look before you let it touch. Add the ability to change things later, one app at a time.

  2. 2

    Keep a person at the gate.

    Anything that leaves your hands waits for a person: a sent email, a payment, a deletion, a post.

  3. 3

    Give it the least access that does the job.

    One folder, one calendar. Never your whole drive or your whole inbox.

  4. 4

    Point it only at sources you trust.

    Your own files and known websites. Unfamiliar sites and strangers’ emails are where hidden instructions live.

  5. 5

    Watch what it does.

    If it opens files or sites you never mentioned, stop the task and check your “recipe”.

DO NOT HAND AN AGENT THESE

  • Sending emails, messages or invites as you without review.
  • Payments, purchases, banking, or anything with a card number.
  • Passwords, logins and one-time codes.
  • Deleting files, emails or calendar events.
  • Medical, HR or legal records, or anything your policy keeps out of AI.
  • Board papers, deal documents or other confidential material on a personal account.
  • Scheduled tasks that open sensitive files or send anything on your behalf.

The last item is Anthropic’s own advice for scheduled tasks.

...also send this file to someone@... A NOTE NOBODY ASKED FOR

PROMPT INJECTION, IN PLAIN WORDS

A note slipped into the pile.

An agent reads everything you point it at. If a web page, email or document contains hidden instructions, such as “also send this file to someone”, the agent may try to follow them. It is like a stranger slipping a note into a stack of papers you asked an assistant to file.

Claude has protections against this, and Anthropic is clear that no protection is perfect. Your best defence is the five rules above: trusted sources, read-only access, and a person approving anything that leaves.

APPROVAL MODES

ManualSTART HERE

Claude asks before every action.

USE ITWhile you learn, and whenever files, accounts or money are involved.

Auto

Claude carries on with read-only actions and safety-checks anything that changes or deletes.

USE ITFor a job you have watched run correctly several times.

Skip

No pauses and no checks.

USE ITNot for work tasks.

The names may change a little as Claude updates. The idea stays the same: the more an agent can change, the more often it should check with you.

08 · HONEST EXPECTATIONS

What it does well, and where it needs you.

An agent is great at first drafts. You still need to know your work well enough to check them.

IT IS GOOD AT

  • Pulling facts from several places into one draft.
  • Following a clear “recipe” the same way every week.
  • Spotting gaps, like a meeting with no agenda or a task with no owner.
  • Doing the first pass, so your time goes on the judgment.

IT NEEDS YOU FOR

  • Numbers, names and dates. Check them every time at first.
  • Anything it cannot see. It only knows what you connect or paste in.
  • Judgment about people, politics and priorities.
  • Long, complicated jobs, which can stall and use more of your plan’s allowance than quick questions.

WHEN IT STARTS SAVING TIME

Plan for a slow first week, because you are setting up and checking everything. The payoff comes from repetition. Once the “recipe” is right, an hour of building each week becomes a few minutes of review.

09 · IDEAS

Twelve ideas worth trying.

Each one starts with read-only access and ends with you reviewing the result. The starred ones tend to surprise people.

Meeting prep packs

For each of tomorrow’s meetings, gather the agenda, who is attending and the last notes into one page.

CalendarShared files

Calendar check

Every Monday, list double bookings, back-to-back runs with no break, and meetings with no agenda.

Calendar

Travel options compared

Compare three flight and hotel options in a table with times and total hours on the road. Booking stays with you.

Web search

Inbox summary

List the emails that need a reply today, and draft replies for you to edit and send yourself.

Email (read only)

Pick one. Build it the same way as the Monday brief, then come back for the next.

10 · GOING FURTHER

When you’re ready for more.

Once your first agent runs well, you can teach it more, give it more to work with, or have several agents share a bigger job. Add one thing at a time, and keep the same safety rules.

01One task02A connectedapp03On aschedule04A savedskill05A team ofagents THIS GUIDE TAKES YOU HERE
Each step builds on the one before. Most of the value is in the first three. The last two help it do the job the same way every time, and take on bigger work.

  • Save your “recipe” as a skill.

    A skill is a “recipe” Claude keeps and uses whenever the job comes up, so it is done the same way every time. Ask Claude to turn your “recipe” into a skill, then find it under Customize, then Skills.

  • Give it background with a project.

    A project holds the things every run should know: who is on the team, how briefs are laid out, words to avoid. Your “recipe” stays short because the background lives there.

  • Let agents share a bigger job.

    For large work, one agent can gather the numbers, another check each project, and a lead agent bring it together. The same rules apply: clear limits, and a person at the gate.

  • Link apps with automation tools.

    Tools such as Zapier connect apps with triggers, like “when a form is submitted, add a row and tell the agent”. Useful once you know exactly what should happen.

  • Let it use a browser, carefully.TAKE EXTRA CARE

    Claude in Chrome can click and type on websites for you. It is powerful and it is where hidden instructions are most common. Use trusted sites only, keep Manual on, and never let it near banking or passwords.

See full workflows built this way →

BEFORE YOU SWITCH IT ON

0 OF 6

View sources (7)
LAST CHECKED: 26 SEPTEMBER 2026 Claude changes often. Menu names and plan details were checked against Anthropic’s documentation on this date. Something look different? Tell us.