Brooke Wright
Brooke Wright · @wright_mode
FREE FIELD GUIDE

Build an AI agent that actually does something.

The visual, no-bullshit guide to the brains, tools, loops and guardrails behind useful AI agents — plus the exact prompt to plan your first one.

✓ 10-minute guide ✓ Copy-paste prompt ✓ No code required

Instant access + a copy in your inbox. No spam.

WRIGHTMODE
THE 10-MINUTE FIELD GUIDE

AI AGENTS
WITHOUT THE
BULLSHIT.

The ridiculously simple guide to understanding, building and actually using AI agents in your business.

NO CODE REQUIRED NO TECH-BRO TRANSLATOR ONE USEFUL AGENT BY THE END
WRIGHT MODE Strategy first. Fancy tools later.
01 / START HERE

THE ENTIRE GUIDE IN ONE SENTENCE

An agent is an AI with a job, access to the tools it needs, and room to decide how to get the job done.

TOO VAGUE “Help me with emails.”
AN ACTUAL JOB “Review new enquiries, work out what each person needs, draft the right reply and flag anything that needs me.”
02 / BUILD THE BORING VERSION

YOUR FIRST AGENT, PLANNED IN 10 MINUTES

Copy this prompt. Give it one annoying job.

Paste this into ChatGPT, Claude or Gemini. It will turn a fuzzy idea into a sane first version — including the bits it should never be allowed to do.

FIRST-AGENT-BUILDER.TXT 
Act as an AI systems strategist. Help me design the smallest useful AI agent for one repetitive job in my business.

The annoying job is: [DESCRIBE IT]

Ask me only the questions you genuinely need to define:
1. The exact outcome the agent owns
2. What triggers it to start
3. The information and tools it needs
4. The decisions it must make
5. What it can do on its own
6. What it must never do
7. Where I need to approve before it acts
8. What context or memory it needs
9. How it checks its work
10. The clear definition of “done”

Then give me:
- a one-page agent brief
- a simple step-by-step process
- the safest first version I can test manually inside this chat
- what I would connect or automate later (only if the manual version works)

Use plain English. Challenge anything vague. Do not add complexity just because it sounds impressive.
TEST 1Run it manually
→
TEST 2Fix the instructions
→
TEST 3Connect tools
→
LATERAdd automation

Do not automate a process you haven’t successfully run by hand. That is how you create a very fast mess.

03 / THE MOVING PARTS

ANATOMY OF AN AGENT

Seven parts.
Zero magic.

The brain, tools and job are the core. These four extra parts make the agent useful without letting it behave like a caffeinated intern.

01

Trigger

What tells it to start.

NEW EMAIL→START
MON 9AM→START
YOU ASK→START
02

Loop

It can check its first answer instead of accepting it like gospel.

DOCHECKFIXREPEAT
“Basically: Hmm. Bit shit. Try again.”
03

Memory

Relevant context from before: customer preferences, decisions, conversations, attempts and business information.

WITHOUT
STARTS AT ZEROWITH
USES CONTEXT
Memory must be intentionally stored or provided. Agents do not all magically “learn”.
04

Guardrails

The permission fence: what it can do, cannot do and must ask about.

CAN ✓CANNOT ×
Useful autonomy. Not an unsupervised toddler with your company credit card.
04 / THE AGENT FORMULA

SCREENSHOT THIS BIT

The whole system,
on one page.

MEMORYwhat it knows
TRIGGER something starts the job
↓
BRAIN looks at the situation
↓
DECISION chooses what to do
↓
TOOLS gets information or takes action
↓
CHECK did that actually work?
↺ LOOP IF NEEDED
↓
DONE / HUMAN APPROVAL
GUARDRAILSwhat it may do
05 / THE LINKEDIN ARGUMENT SHORTCUT

IS IT ACTUALLY AN AGENT?

A deeply scientific decision tree.*

DOES IT HAVE A JOB?
NO → probably not an agent. YES ↓
CAN IT DECIDE HOW TO COMPLETE THAT JOB?
NO → probably an automation or workflow. YES ↓
CAN IT USE TOOLS OR TAKE ACTION?
NO → useful assistant, maybe. YES → NOW WE’RE GETTING AGENTIC.

*“Agent” is not a perfectly agreed technical category. Understand the useful architecture. Leave the terminology cage fight to LinkedIn.

06 / PICK THE SIMPLEST SYSTEM

AUTOMATION VS AGENT

Not everything needs a bloody agent.

PREDICTABLE

AUTOMATION

IF X → DO Y

Best when the steps and answer are known every time.

EXAMPLESomeone fills in a form → add them to a list.
  • Fast
  • Cheap
  • Consistent
  • Easy to test
VS
ADAPTABLE

AGENT

GOAL → LOOK → DECIDE → ACT

Best when the situation changes and judgement is genuinely required.

EXAMPLERead an enquiry → understand the need → choose the service → prepare the next action.
  • Flexible
  • Handles messy inputs
  • Can use context
  • Needs tighter guardrails
Choose the simplest system capable of doing the job.
07 / AUTONOMY IS A DIAL

HUMAN APPROVAL

You do not have to choose between “AI does nothing” and “AI runs the company”.

LOWER RISK

AI SUGGESTS

↓

HUMAN APPROVES

↓

AI ACTS

Use this for new agents, customer communication, money, legal commitments and anything hard to undo.

MORE AUTONOMY

AI ACTS

↓

HUMAN REVIEWS

↓

EXCEPTIONS GET FLAGGED

Use this after the agent has earned trust on a narrow, reversible job with excellent logs.

Start with approval. Remove approval steps only when the evidence says you can — not because the demo looked cool.
08 / PERMISSION FENCE

GUARDRAILS THAT ARE ACTUALLY USEFUL

Write the rules before you hand over the keys.

CAN DO

✓ Read new enquiries

✓ Categorise them

✓ Draft replies

✓ Update the CRM

✓ Flag exceptions

CANNOT DO

× Send without approval

× Issue refunds

× Offer discounts

× Delete records

× Invent missing facts

WHAT CAN IT READ?WHAT CAN IT CHANGE?WHAT NEEDS APPROVAL?WHAT IS NEVER ALLOWED?WHAT SHOULD TRIGGER A STOP?WHAT GETS LOGGED?

A useful guardrail is specific. “Be careful” is not a guardrail. It is a hope wearing a hi-vis vest.

09 / REAL AGENT #1
01

INBOX AGENT

Keep the owner’s inbox under control.

TRIGGEREvery weekday morning
TOOLSGmail + business context
APPROVALOwner approves every reply
READ→CATEGORISE→DECIDE→DRAFT→FLAG→APPROVE
INBOX-AGENT.TXT
You are my inbox triage agent. Your job is to help me start each weekday knowing exactly what needs attention.

For every unread email since your last run:
1. Categorise it: urgent, needs reply, task, FYI, newsletter or spam.
2. Identify what the sender needs and any deadline.
3. Draft a reply in my supplied voice when a reply is needed.
4. Flag missing information, money, complaints, contracts or sensitive decisions.
5. Finish with a short priority list for me.

Never send, delete, agree to pricing, change a contract or promise a deadline. If context is missing, ask instead of guessing. I approve every reply.
10 / REAL AGENT #2
02

CONTENT REPURPOSING AGENT

Turn one real idea into useful content drafts.

TRIGGERNew transcript or video
TOOLSTranscript + voice guide + content library
LOOPDraft → critique → rewrite
READ→FIND IDEAS→CREATE→CRITIQUE↺REWRITE→APPROVE
CONTENT-AGENT.TXT
You are my content repurposing agent. Turn the supplied source into strong drafts without inventing ideas I did not say.

1. Read the complete source before writing.
2. Identify the 5 strongest original ideas, stories or opinions.
3. Choose the best format for each idea and explain the choice in one line.
4. Draft the content using my brand voice guide.
5. Critique every draft against the source, voice rules and banned-phrase list.
6. Rewrite anything generic, repetitive or unsupported.

Keep my original meaning and strongest wording. Flag claims that need checking. Do not publish. Finish with the drafts ranked from strongest to weakest for my approval.
11 / REAL AGENT #3
03

LEAD RESEARCH AGENT

Walk into a sales call already understanding the room.

TRIGGERNew qualified lead
TOOLSWebsite + web research + CRM
GUARDRAILNever contact the prospect
RESEARCH→MODEL→PROBLEMS→OPPORTUNITIES→QUESTIONS
LEAD-RESEARCH-AGENT.TXT
You are my pre-call research agent. Prepare a factual briefing for this qualified lead using their website, reputable public sources and the CRM notes I provide.

Find:
- what the business sells, to whom and how it appears to make money
- current offers, team and recent relevant activity
- likely operational or growth problems, clearly labelled as hypotheses
- where my services may be relevant
- 7 specific questions that will help me test those hypotheses on the call

Link every source. Separate confirmed facts from inference. Do not guess revenue, scrape private data or contact the prospect. Finish with a one-screen call brief, not a biography.
12 / MULTI-AGENT SYSTEMS

YES, THEY CAN WORK TOGETHER

Three specialists.
One shared outcome.

01

RESEARCH AGENT

Finds reliable source material.

↓
02

WRITING AGENT

Turns the research into a draft.

↓
03

REVIEW AGENT

Checks accuracy, voice and rules.

DO NOT BUILD THIS FIRST.

Start with one agent doing one useful job. A team of agents is still a team you have to design, supervise and debug. Complexity is not the goal.

13 / TEAR-OUT WORKSHEET

BUILD YOUR OWN AGENT

Make the job clear before you touch the tools.

14 / YOUR FIRST AGENT

THE SEVEN-QUESTION CHALLENGE

Find one repetitive job that makes you say, “Oh FFS, not this again.”

01What outcome do I want?
02What information does AI need?
03What tools does it need?
04What decisions does it need to make?
05What should it NEVER do?
06Where should a human approve?
07How will it know it did a good job?

YOUR FIRST AGENT SHOULD BE BORING.

It should solve one annoying problem ridiculously well.

You can make it fancy later.
15 / THE BIGGER PLAY

YOU UNDERSTAND AGENTS. GOOD.

Now build a business where AI does more than write the occasional caption.

Understanding agents is step one. Building an AI-first business is the bigger play.

Inside Wright Mode, I teach business owners how to spot the right jobs, design sane workflows and use AI without handing their judgement to a robot.

LEARN WITH WRIGHT MODE →