Give It a Goal

I have a habit of asking AI for one small thing, then discovering that what I actually wanted was an entire application.

The first request is usually reasonable. The next one begins, “Could you also…”

“Could you also” is doing a lot of work in my life.

Looking at how I use AI, I can see six different ways of handing over work. I still use all of them. Asking for a working prototype hasn't stopped me needing spelling help in the next message.

1. Google, but I can interrupt

At its simplest, I use ChatGPT or Claude to find something out. Explain this technology. Compare these approaches. Help me understand the subject before I develop a strong opinion on it.

The useful part is being able to say, “I still don't get it.” I can keep asking without anyone developing an increasingly strained smile.

I still need to check important facts and open the sources. A patient explanation can be wrong too.

2. Help me sound like myself

Then there's writing: fix the spelling, shorten this paragraph, help me get to the point.

A typical request is, “Make this clearer and warmer, but keep what I'm trying to say.” If it comes back sounding like a company announcing a merger, we have another go.

There's something funny about asking a machine to help me sound more human. But it helps. I bring the thought, it helps with the wording, and I decide whether I'd actually send it.

3. Here. You do this bit

The next step is handing over a small job.

For a spreadsheet, that might be: check the figures, trace anything suspicious, and give me a list of proposed corrections. Keep the original intact. Show me what you found.

With the right tools, an AI agent can work through those steps. I get something to review without opening every tab myself.

“Find the errors” and “change the file” are separate permissions. Being clear about that lets me hand over the tedious part without wondering what else has changed.

4. Make me something I can use

Now the request becomes a whole result: a report, a website, a prototype.

I've used AI to turn rough ideas into things I can actually try. Once there's something on screen, I can see what's missing, what feels wrong, and what was more complicated in my head than it needed to be.

Apparently I have very specific opinions about things I described as “something simple”.

That's useful discovery. It also means I have to check the result: does the website work on a phone? Do the report's sources support its conclusions? What happens when someone clicks the wrong thing?

“Looks finished” has caught out enough humans without us automating it.

5. Get a team on it

For larger jobs, I've started asking for separate agents to research, design, build and review. A coordinating agent can divide the work and bring it together—one practical way to organise an agent team.

Some jobs can happen at the same time. Another agent can check the result against the brief.

But calling an agent “Chief Quality Officer” doesn't make it better at checking a link. It needs a clear test and evidence that it passed.

Otherwise I've recreated a meeting where everyone says it looks good. I already know how to have that meeting.

6. Let the team take shape around the work

This is where my experiments are heading: could the system work out what help a goal needs, divide the work, and adjust as it learns?

That's what interests me about agent swarms. The terminology is loose, but I think of a team as having fairly clear roles. A swarm is more fluid: agents explore different paths, hand work around and bring in help as needed. The two overlap.

That could suit a broad investigation or a problem with lots of independent parts. It also needs limits. More agents can mean duplicated work, shared mistakes and a larger bill.

I'd like to know when it will stop before I go and make a coffee.

The bigger change: give it the goal

Across all six, the thing I'm learning to do better is explain what I'm trying to achieve.

I can ask for a report when I need help making a decision. Or ask for a website without saying what a visitor should understand or do.

The AI gets the task. The useful context stays in my head, where access is limited and the documentation is poor.

Task prompting says: “Build me a website.”

Goal prompting gives it a reason for the work:

Help a first-time visitor understand this service and decide whether to get in touch. Build a simple preview using the approved information. Make it work on a phone, check the links, and flag anything missing. Work for up to an hour, then stop for my review. Don't publish it.

Now it knows the purpose, what to check, and where to stop. It has room to work out the steps.

This helps with a single email too: “Help the reader understand the delay and what happens next. Keep it brief. Don't invent a date to make the ending more cheerful.”

No swarm required. Just a clearer explanation of what would help.

I still move between all these levels. Some jobs need a quick answer. Some deserve a team. The useful skill is choosing how much to hand over, giving enough context, and checking that the result does the job.

I'm getting better at that.

And, occasionally, at recognising when “could you also” should become “that'll do, send the email”.