The job description nobody writes

Thanks for coming back for issue five. This fortnight one pattern showed across several completely unrelated sessions: the businesses getting flat, generic output from AI aren’t hitting a limit in the model. They’re skipping a step that turns an AI tool into a useful collaborator, and it isn’t a technical step.

The pattern: no job description, so you get a first-year graduate

The owner of a multi-business group described his experience with AI to me plainly: “I just treat this like having a first year graduate who I’ve got to say, no, look, this needs to be done. And that’s my experience coming through, like, for quality stuff. And it’s giving me back first year graduate type answers.”

When I looked at his setup, the reason was obvious. He had projects created for each of his businesses, but every one had an empty instructions field. He was doing the equivalent of hiring someone competent, sitting them at a desk, and never telling them what the role actually is, what “good” looks like, or how the business operates. Then judging them on the output.

That’s not a one-off. I’ve now seen the identical gap in businesses ranging from a professional services firm to a private investment office: capable tools sitting mostly idle or under-used, because nobody has told them what job they’re doing.

Why the fix is a job description, not a better prompt

The instinct is to write a better one-off prompt. That helps a single answer. It doesn’t fix the underlying problem, because a new chat window has no memory of anything you told it last time.

The better fix is treating a persistent AI workspace, a Project, in the tools most people use, exactly like you’d onboard a new employee. Not a task list. A job description: who this is for, what the role covers, the tone and format expected back, and what questions it should ask before it starts rather than guessing.

I use this on my own work constantly. I run a project for my kids’ swim coaching, and its very first instruction is a clarifying question, not a task: before doing anything, check which pool it is. 25 metres or 50, because the training set changes completely depending on the answer. The AI won’t proceed until it knows. That one line of instruction does more for output quality than any amount of prompt-polishing on the actual question.

Once you write that brief (good instructions) once, every conversation inside that project inherits it. The tool stops behaving like a search engine you interrogate from scratch each time, and starts behaving like someone who already knows the business.

What this actually looks like

For a business owner or executive reading this on a Sunday night wondering why AI hasn’t clicked yet, the fix is a twenty-minute job, not a training program:

• Open the instructions field (look for the three little dots when inside a project, often the top left of the screen) on whatever persistent workspace your tool offers (Claude Projects, ChatGPT Projects, or equivalent), most people have never touched it

• Write it as a job description, not a task: who is this for, what does the role cover, what tone and format should it default to

• Add one instruction telling it to ask a clarifying question before responding, rather than guessing at what you meant

• Give it the two or three documents (look for context section inside the project) it actually needs to know your business, values, a scorecard, a style guide, whatever it is you’d hand a new hire in week one

None of this changes what the model is capable of. It changes whether the model knows what it’s meant to be doing for you. That’s the gap, in most businesses I sit down with, not a capability problem, a briefing problem.

Have a go at this this week and compare the results you get.