Part 1 of 4 | Build and test your first AI job
This guide will help you choose one useful task, test it in a fresh Claude chat, and save the result as a simple checklist you can reuse.
If you are new to AI, that is fine. You do not need to know any special terms. The guide will tell you what to type, what you should see, and when to stop for review.
Before you start
What you need
A computer with internet access
A Claude account
One real business goal or recurring frustration in mind
A folder where you can save simple text files
What you do not need
Coding experience
A terminal or command line
Claude Code yet
Automation software
What this guide will not ask you to do
Send anything to a customer
Publish anything
Change live systems
Spend money
This is a draft. You stay in control.
What you will produce
By the end, you should have:
one real task chosen
one reason it is worth testing
one short job brief
one draft result from Claude
one small checklist to reuse next time
All that in about 45–60 minutes.
Why this matters
A good first AI job should be useful, but it should also be safe to learn from.
You are looking for a task that is:
Valuable: it helps with a current goal or a repeated problem.
Ready: you can provide the needed information and judge whether the result is good.
The goal is not to make your first job autonomous. The goal is to make it clear, testable, and reusable.
Phase 1: Choose one real goal and list possible tasks
Time: 10 minutes
Step 1: Write one real goal
Write one sentence:
Over the next 90 days, I want to…
Examples:
Increase qualified sales calls.
Reduce the time spent on weekly reporting.
Improve the quality of onboarding for new hires.
Understand why customers are leaving.
Prepare a cleaner monthly review for the leadership team.
Stop and check: If your sentence is broad, make it narrower.
Too broad: Grow the business.
Better: Increase the number of qualified sales calls over the next 90 days.
Step 2: List five possible tasks
Write down five jobs that support the goal.
Examples:
Researching prospects
Summarising customer calls
Drafting follow-up emails
Comparing weekly performance numbers
Preparing a meeting brief
Do not try to choose the perfect one yet.
If you get stuck, open a fresh Claude chat and type:
My goal is: [write your goal].
I run: [describe your business in one sentence].
List 10 practical tasks that support this goal.
Focus on tasks involving reading, comparing, summarising, drafting, checking, or organising information.
Do not recommend sending, publishing, purchasing, deleting, or changing live systems.
AI will later turn this into a draft job. For now, you are only making a shortlist.
Step 3: Ask the practical question
Look at your five tasks and ask:
Which one wastes time, causes delays, or keeps happening?
Put a star next to two or three candidates.
What to do now
Reply to this email and tell us:
Which task would you like to test first?
Phase 2: Pick the best first task
Time: 10 minutes
Some tasks matter a lot but are poor first tests.
Examples:
Approving a large payment
Sending a message to every customer
Changing prices
Hiring or firing someone
Running payroll
Those may become AI-assisted later, but they are not ideal first experiments.
For your first test, choose something that produces a draft for review.
Now that you have examples, let’s pick the best first task.
Step 1: Draw a simple 2×2 matrix
On paper or in a document, draw this:
Low Readiness | High Readiness | |
|---|---|---|
High Value | Important, but not ready | Best first task |
Low Value | Ignore for now | Quick win only |
If you prefer plain English, use these four boxes:
Best first task: useful and easy to judge
Important, but not ready: useful, but missing information or too risky
Quick win only: easy, but not important
Ignore for now: low value and low readiness
Step 2: Score each candidate
Give each task two scores from 1 to 5.
Score | What to ask |
|---|---|
Value | Would this help a current goal or deal with work that keeps coming back? |
Readiness | Can I provide the information and judge the result? |
Overall | How do Value and Readiness combine in this recommendation? |
Need help? Use these examples.
Weekly KPI summary: Value 4, Readiness 5
Drafting customer follow-ups: Value 4, Readiness 4
Changing product pricing: Value 5, Readiness 1
Renaming old files: Value 1, Readiness 5
A task can be important but still be a bad test for today.
Step 3: Choose your winner
Choose the task with the best mix of value and readiness.
The right balance is enough value to matter and enough readiness to test safely.
If there is a tie, choose the one you can test this week using information you already have.
If you are unsure, use AI to sanity-check the choice
Type:
Here are three tasks I am considering:
1. [task]
2. [task]
3. [task]
Score each one from 1 to 5 for:
- Value: how much it supports my current goal or reduces repeated friction
- Readiness: whether I can provide the needed information and judge the result
Recommend the best first test.
Explain your reasoning in plain English.
Do not recommend any task that requires sending, publishing, purchasing, deleting, or changing live systems.
Use the matrix to place your tasks by Value and Readiness. Choose the one that gives you enough value to matter and enough readiness to test safely.
What to write on each note
Write the same three lines:
Task: one short job name
Value: 1-5
Readiness: 1-5
Sample card
Task: Weekly KPI summary
Value: 4
Readiness: 5
What the scores mean
Use Value and Readiness as the input. Then place the task in the box that best matches those scores.
The box gives you the recommendation:
High Value + High Readiness: Best first task
High Value + Low Readiness: Important, but not ready
Low Value + High Readiness: Quick win only
Low Value + Low Readiness: Ignore for now
These scores produce one final recommendation. They are not a third scoring system.
Phase 3: Turn the task into a clear job brief
Time: 15 minutes
A good AI job has a clear result, clear inputs, clear checks, and clear boundaries.
You will write six lines:
Outcome: What should exist at the end?
Inputs: What information should AI use?
Allowed sources: Which files, pages, or notes may it rely on?
Checks: What should it verify before saying it is done?
Stop points: When must it ask you instead of continuing?
Evidence: What should it show you so you can verify the result?
Example job brief
Task: Prepare a weekly KPI summary
Outcome: A one-page summary of weekly sales, leads, and conversion rate
Inputs: This week’s report and last week’s report
Allowed sources: Only the two reports I provide
Checks: Show the calculation for every percentage change
Stop points: Ask me if any figure is missing or unclear
Evidence: Show the source numbers and the final comparison table
Copy this template
Task:
Outcome:
Inputs:
Allowed sources:
Checks:
Stop points:
Evidence:
Use AI to improve your brief
In a fresh Claude chat, type:
I want to test this task:
[paste your task]
Ask me up to five questions to turn it into a clear job brief.
The final brief must include:
- Outcome
- Inputs
- Allowed sources
- Checks
- Stop points
- Evidence
Keep the task draft-only.
Do not send, publish, purchase, delete, or change any live system.
Answer the questions.
You’ll get a job brief you can review before any work starts.
Where to type this
For Issue #1, use a normal Claude chat.
Open Claude in your browser.
Start a new chat.
Paste the task-refinement prompt.
Answer the questions.
When Claude shows the final brief, review it before moving on.
Need an example? See this below:
After Claude gives you the brief, type:
Show me the brief again as a clean checklist. Do not start the task yet.
Phase 1: Ask AI to turn the brief into a reusable mini-checklist
Once the job brief looks right, ask AI to create a reusable checklist.
Type:
Turn this approved job brief into a reusable checklist.
Use these sections:
- Purpose
- Inputs needed
- Steps
- Checks
- Stop points
- Evidence to save
Keep it short and practical.
Do not add permissions I did not give you.
Save the result as:
task-name-skill.md
Example:
weekly-kpi-summary-skill.md
A simple text file is enough.
What this checklist does
It records:
What the task is for
What information is needed
What steps to follow
What to check
When to stop
What proof to save
This is the first step toward reusable AI work.
In later issues, this becomes part of your AI starter kit.
Optional side quest: Prepare your local AI starter kit
Time: 10 minutes
This is optional for Issue #1.
If you want to organise your work now, download and unzip the starter folder provided with this issue.
What is inside the starter
File or folder | What it is for |
|---|---|
START_HERE.md | Starting instructions for the series. |
AGENTS.md | Working rules, including when AI should stop for review. |
Reference | Business information you confirm. |
Skills | Saved ways of doing jobs. |
Projects and Outputs | Current work and useful results. |
Runs | Records of what AI used and checked. |
Do not worry if some folders are empty.
If you skip this today
Save your job brief and checklist somewhere easy to find.
You can place them into the starter folder in a later issue.
Next week
In Issue #2, we will install Claude Code and use it inside the starter folder.
How to keep the task safe
For your first AI job, keep the result as a draft.
Use this rule:
Draft first. Review second. Act only after approval.
Good stop points include:
Ask me if required information is missing.
Do not use sources outside the ones I provide.
Do not publish or send anything.
Do not change live systems.
Do not spend money.
Do not delete files.
Do not include sensitive personal or customer data unless I explicitly provide it for this task and confirm it is allowed.
Review alignment before the first run
Stop here for two minutes.
Before you run anything, confirm:
Is this the right task?
Is the desired outcome clear?
Are the allowed sources clear?
Are the stop points clear?
Is the result still draft-only?
If another person cares about the result, ask them to confirm the job brief before you continue.
Audit fit
Your first run should leave a small evidence record.
Create a note called:
RUN_LOG.md
Record:
Task name
Date
Sources used
Checks performed
Stop points triggered
Draft output saved
Final decision
Simple evidence record example
Task: Weekly KPI summary
Date: 2026-09-10
Sources used: weekly-sales.csv, previous-week.csv
Checks performed: totals verified, percentage changes recalculated
Stop points triggered: none
Draft output saved: weekly-kpi-summary-draft.md
Final decision: Approved for internal use only
You do not need a fancy audit system yet.
You just need enough evidence to answer:
What did AI use, what did it check, and what did I decide?
Phase 4: Run the first test
Time: 15 minutes
Now run the task using a small, safe set of information.
Use a small test
Good:
One customer call transcript
One week of numbers
Five prospect records
One meeting’s notes
Not yet:
The whole customer database
Every company file
A live workflow
Anything that sends or publishes automatically
Test-run prompt
Start a fresh Claude chat and paste:
Use the approved job brief below to complete one small draft-only test.
Job brief:
[paste the brief]
Test inputs:
[paste or attach the small test inputs]
Rules:
- Use only the inputs I provide.
- Show your checks.
- Stop and ask me if information is missing or unclear.
- Do not send, publish, purchase, delete, or change any live system.
- Keep the result as a draft.
Review the result
Check:
Did it use only the allowed information?
Did it produce the requested outcome?
Did it show the checks?
Did it stop where it should?
Is the evidence clear?
Now compare the draft to what “good” looks like for you. A technically correct result can still be unhelpful if it does not match your real goal.
Decide what to do next
Choose one:
Keep: good enough to reuse
Improve: useful, but the brief or checklist needs changes
Stop: wrong task, poor fit, or not safe enough
What to save
Save these three files:
Your job brief
Your reusable checklist
Your evidence note
Suggested names:
weekly-kpi-summary-brief.mdweekly-kpi-summary-skill.mdRUN_LOG.md
What counts as evidence?
Evidence is what lets you verify the result without trusting a claim.
Examples:
The source files used
The calculations shown
A comparison table
A list of missing information
The saved draft file
If AI says “done,” but you cannot inspect the proof, the task is not finished.
Audit your evidence
Time: 5 minutes
Think of evidence as proof that you can inspect later.
Ask AI to add an evidence section to every test output.
Type:
Add an evidence section that lists:
- Files or text used
- Important assumptions
- Checks performed
- Missing information
- Draft output created
- Questions that still need my decision
Read the evidence section before you approve the result.
When to stop and ask for help
Stop the task if:
You cannot tell whether the output is correct.
The task needs information you should not share.
The result would affect customers, staff, payments, or legal obligations.
The AI wants to use a source you did not approve.
The task starts expanding beyond the brief.
The output would be hard to undo.
Stopping is not failure.
It is part of the design.
If the task does not work
Choose the smallest fix.
If the output is:
Too vague: add an example of the desired result.
Incorrect: add a check.
Using the wrong information: narrow the allowed sources.
Too risky: add an earlier stop point.
Too large: reduce the test inputs.
If the job itself was a poor choice, return to the matrix and choose another candidate.
There will usually be a better test than forcing the wrong one.
How to improve the brief
Add one sentence at a time.
Examples:
Use only the attached files.
Show every calculation.
If a number is missing, stop and ask me.
Keep the result under one page.
Do not contact anyone.
These small instructions make the job more reliable.
Define “done” clearly
Finish this sentence:
This task is complete when…
Example:
This task is complete when the draft summary includes this week’s numbers, last week’s numbers, percentage changes, and a list of any missing data.
This sentence gives AI the outcome you expect and gives you a final check.
Expected outcomes
By the end of Issue #1, you should have:
One real business goal
A short list of candidate AI jobs
A Value and Readiness score for each
One selected first task
A written job brief
A reusable checklist
One draft test result
One evidence note
A clear Keep, Improve, or Stop decision
What you learned
You learned that useful AI work begins before the prompt.
It begins with:
Choosing the right task
Defining the desired result
Naming the information AI may use
Adding checks
Adding stop points
Saving evidence
You also created the first reusable “way of working” that can later become part of your local AI system.
What changed
Before: “I should probably use AI for something.”
After: “I have one useful task, a clear brief, a repeatable checklist, and a verified draft result.”
What this prepares you for
Next, you can move from one-off chats to a local workspace where:
Rules are saved
Reference material is separated from drafts
Skills can be reused
Runs can be logged
Work can stay draft-only until approval
Reflection
Ask yourself:
Was the chosen task more useful than expected?
Which instruction made the biggest difference?
Where did AI need clarification?
What check caught a mistake or gave confidence?
Would you reuse this checklist?
Write one sentence:
The main thing I would change next time is…
Final checklist
I chose one real goal.
I listed at least five candidate tasks.
I scored the best candidates for Value and Readiness.
I chose one safe first task.
I wrote a six-part job brief.
I had AI turn it into a reusable checklist.
I ran one small draft-only test.
I reviewed the result.
I saved evidence.
I decided to Keep, Improve, or Stop.
Next week
In Issue #2, you will install Claude Code and open your AI starter kit on your own computer.
You will not build a complex system yet.
You will simply learn how to open the workspace, read the rules, and run a local draft-only task.
