Artificial Intelligence

Build an AI That Writes Your Tasks For You

In my last article I shared the workflow of the five field task brief I use every day: TitleGoalTools and WorkflowStepsAcceptance Criteria The framework works. ...

infoguy 5 min read · 2 months ago

5 2

In my last article I shared the workflow of the five field task brief I use every day:

Title
Goal
Tools and Workflow
Steps
Acceptance Criteria

The framework works. But writing out those instructions by hand after every client call costs time, and that is exactly the time we are trying to save.

It also lands at the worst possible moment, when your brain is already drained from listening, thinking, and deciding. That is when you are trying to remember the details, structure them clearly, and not miss anything.

So here is the next step. You set the framework up once inside an AI. After that you paste in the raw client message or the call transcript, and you get a finished, structured task back.

The Platform Barely Matters

You can build this as any of the following:

A Skill in Claude
A Custom GPT in ChatGPT
A Gem in Gemini

For what we are doing here, people spend too long on this decision. All three hold a set of instructions permanently so you never type them again, and that is the whole job at this stage. Use whichever one you already work in.

Two practical differences worth knowing before you start:

Instruction length. Gems keep you to a short instruction box and Custom GPTs give you more room, while Claude Skills have no real limit. If your rules get long, that matters.
How they load. A Custom GPT or a Gem is a place you go to. A Claude Skill loads itself when the work matches, which becomes important later.

What matters right now is what you put inside it.

Setting It Up

You do not have to write the instructions yourself. Here is the process:

Open the AI and tell it what you do. Your business, the services you provide, what your projects usually look like, and how your workflow runs.
Paste in the five field framework.
Ask it to turn all of that into a set of instructions for a skill that takes raw input and returns a structured task.
Read what it gives you and fix anything that does not match how you actually work.
Save it.

From then on the job is simple. Paste in the meeting notes, the call transcript, or the scattered client message, and the structured task comes back.

This is the single skill version, and it is where everyone should start. There is a bigger setup where you run several skills that feed each other, one holding your domain knowledge and another holding your company policies, but that deserves its own article and I will cover it next. First get this one working.

build-an-ai-that-writes-your-tasks-for-you

Rules That Save You From Checking

You can also write conditional rules into your instructions, and this is what turns a template into a system.

Say your team monitors reviews on an ecommerce store you manage. You can build the rule into how those tasks get written:

If there are negative reviews, post a comment in the task activity with the screenshot and nothing else.
If there are no negative reviews, post nothing at all.
If a specific error shows up, tag a specific person in the activity.

The point is not the rule itself. The point is that once it is written into the skill, you stop checking whether it happened. You are not opening the task to see how the comments were handled or whether anyone was told. It is already handled the same way every time.

That is the real saving. Not the minutes spent writing the task, but the attention you get back from not having to verify things.

What Went Wrong For Me

My first version was generic, and I want to show you the specific failure that taught me the most.

I pasted in a client message that covered several different things at once. The AI took all of it and produced one task. It looked thorough. The steps covered frontend work, backend work, and monitoring the sales reports, and every point in the acceptance criteria was reasonable on its own.

But nobody could complete it:

The frontend person could not finish it, because half of it was not their work.
The backend person had the same problem.
The acceptance criteria were too long to check, so nobody could say when it was done.

The task was not too big. It was three tasks wearing one title.

One Task, One Owner

That failure gave me the rule I use most. A task should belong to one person. If a brief spans work that different people would do, it is not one task, it is several.

There is also a signal that tells you when you have got it wrong:

If the acceptance criteria are too long to check quickly, you have merged tasks that should have stayed separate.

Long acceptance criteria are almost never a sign of thoroughness. They are a sign that something got combined that should not have been.

There are times when a piece of work genuinely needs more than one person, and splitting those cleanly gets detailed fast. That is one of the areas I am working through properly in my first book, with the case studies behind it. But for most of what you assign in a week, one task one owner will hold, and it removes most of the confusion before it starts.

How The Tuning Actually Happens

This is worth saying plainly, because people expect the setup to be right on day one. Mine was not.

The method is simple and it never really stops:

Watch what comes out.
Find where it breaks.
Write the fix into the instructions.

Over time the skill stops repeating the mistakes you have already caught, and the instructions get sharper every time you do it.

The same goes for input that arrives incomplete, which happens constantly with client messages. There are ways to make the skill handle that instead of guessing, and it is a longer topic than it looks, so I am covering it properly in my first book rather than in a paragraph here.

Start This Week

Build the first version this week. Tell it what you do, give it the five fields, and start pasting in your raw notes.

It will not be perfect. Mine was not. The first few tasks it gives you will show you exactly which instruction is missing, and that is the point. The tuning is what makes it yours.

Whether you are a freelancer writing tasks for your own future self, a contractor coordinating a small team, or running a startup, the setup is the same. Only the context changes.

What is the first thing you would put in your instruction set?