Part 2: Show your AI what a good result looks like

Part 2 of 4 | Build and test your first AI job

Imagine you're about to hand your client-call recaps to AI.

You give it a transcript and ask for a summary.

Without an example or a clear explanation of what you want, AI has to work out how much detail to include and what belongs in the client version.

Since the AI has less guidance, the draft might have some mistakes in its output.

You end up rearranging the recap, checking promises against the transcript, and explaining what to leave out.

The writing is done, but you're still working out how to make it usable.

But having the example in front of you makes it easier to point out what works.

You open one output you approved to show it what you want.

You'll turn your idea of good work into instructions AI can follow.

So prepare an example you like, or get ready to talk through the result you want.

AI will organize the details and ask about anything it still needs you to decide.

You already chose the job in Part 1. Now you're giving it your standard, so you have something specific to use in judging the first result.

Use the same AI Brain Starter project from Part 1.

The more specific the instruction to AI is, the better it does the job.

That's why it needs something specific that would keep it from guessing. Something like a job card.

Remember the job summary you approved in Part 1?

The job card is its expanded, more detailed version.

It tells your AI Brain what to produce, which inputs to use, and how to check the result.

It also tells you what you need to provide and which decisions still belong to you.

Open the same project chat and send this prompt:

Checking the plan now gives you a chance to catch any inaccuracies before AI builds a result around it.

The response you got from AI likely looks something like this:

Let's go through it part by part.

This is a quick reiteration of the job.

Just check that the AI describes it accurately.

The succeeding sections need a closer look.

Like what I got here:

How much AI asks for depends on the business information already in your initial business files.

It may have a useful example from it, or it may still need you to show it what a good result looks like.

The complete card should make the starting point and expected result clear.

Check the “Ready-to-run job card”.

Here's another look at the one I shared above:

Check that:

Most of the time, it's hard to explain your standards for the quality of your output.

It's better to just show it.

Supply whatever it is that helps define the job.

Attach the file to the chat, or paste the relevant section, or add a readable screenshot.

If the sample is already in the initial files you uploaded, point AI to it. Explain what to learn from it and which details don't apply.

⚠️ Before sharing an example

Remove passwords, payment details, personal data, and confidential information you aren't allowed to share with the AI service.

You can still keep going.

Tell AI you don't have an example and ask it to show you your options:

The next step depends on what's missing and what's already saved in your project.

AI may ask for a smaller real example, such as one useful section of a report.

It may also suggest rules for tone, format, or wording based on material you've already supplied.

Depending on the case, the AI may suggest options and rules for the job's output. Especially if you don't have a sample to give.

Like it does here:

Check the source it used, correct anything that doesn't fit, and ask for a revised version.

Your sample or answers may change more than one part of the plan.

Ask AI to show the complete updated card so you can check how the instructions fit together.

Check each revision every time.

Explicitly telling the AI to save the job card gives the first run your agreed instructions, including any changes from the samples and conversation.

It also serves as the universal instruction for every time you need to do that job in the future.

When the card matches what you want and you've settled the decisions needed for the first test, say:

This will save the job card into a file that Codex can read for future sessions.

Notice that “approved job card” is in blue.

That's because it's clickable (it will show you that saved job card).

Open what AI saved. Check that it is updated with your corrections and matches what you want.

If you're happy with the saved version, you've finished Issue 2.

If the file is missing or wrong, check the Optional Help below before continuing.

You now have a specific job to hand off to AI and a standard for the work.

In Part 3, you'll get the first real result and judge it against that standard.

See you there.

Continue to Part 3

Choose the easiest way to supply the detail AI needs. Apply the privacy check before sharing anything new.

For example:

A project schedule can give useful context, but a scheduled date doesn't establish that someone agreed to it on a call.

If sources disagree, ask AI to flag the difference for your decision.

Choose something close to the result you want from this job.

A short recap you actually approved can teach more about your preferences than a polished report with a different purpose.

You can use part of a document.

Point out what you like and what needs changing.

If it takes several examples to explain the result, give each one a role: one might show the format, while another shows the level of detail.

You will eventually establish and test possible routes for another case, including where input comes from and where you review the result.

But for now, let's keep it simple.

A file, export, or pasted passage is enough if it gives you what you need to define the job.

The card can name an input you'll supply later for the run. Resolve anything that prevents agreement on the result or its checks before approving it.

Ask AI to fix the saved version:

Open the file again and check your corrections.

If AI reports an access problem, resolve it before continuing.

A new chat needs access to the same project; point it to the file you checked if it can't find the accepted card.