Since I started using generative AI at work, I have come to spend more time adjusting the setup around the models than working with the models themselves.
What information should I provide? How far back should the AI look? What rules should it follow, and which tools should it have access to? How much should it be allowed to decide on its own?
This is the work of adjusting the environment in which AI operates—what is often called a “harness.”
Lately, I have been using one measure to see how well that setup works.
When I say “make it good,” how good does the result actually get?
I think this is one of the things that distinguishes how effectively people use generative AI today.
Of course, “make it good” is not a particularly brilliant prompt. As an instruction, it is pretty vague.
And yet, it produces good results. How far have I managed to build an environment in which those few words are enough to get the work done? That is what interests me.
What can AI actually see of a company?
Try asking the AI you use at work:
“What kind of company are we?”
If it can only access the company website, its answer will probably center on the company profile and service descriptions. If it has only been given documents for a particular project, that project will define its view of the company.
But doing the actual work requires more than that.
What are our main products, and who buys them? What problems do those customers tend to have? What have we proposed to them before, and what won us the business? Conversely, which kinds of deals have we lost?
What the sales team is thinking. The constraints the technical team understands. The areas management wants to develop next.
The company’s decisions rest on the accumulation of all this information.
Yet it is scattered across Google Drive, email, chat, and people’s heads. Even when a document has been saved, it cannot inform the work if the AI cannot access it and find it as relevant information.
There is still a gap between a company having information and AI being able to use it.
What lies behind “make it good”
People give each other instructions like these all the time:
“Like last time.”
“Make it good for this customer.”
“Put it in the usual format.”
The person who understands these instructions knows what “last time” and “usual” mean. They have worked alongside you, know the company’s circumstances, and remember both the customer’s preferences and the proposals that did not work out.
Behind the few words “make it good” lies a great deal of shared context.
But even people do not always understand each other.
The person asking might mean “finish it quickly,” while the person doing the work hears “polish every detail.” Even within the same company, with the same documents in front of them, different priorities can lead to different results.
Information is not the only thing that needs to be shared. So do the criteria for deciding what counts as “good.”
I think this matters just as much when working with generative AI.
Suppose every request requires a long prompt, starting with an explanation of the company, then the customer’s circumstances and the history of the work. Alongside the wording of the instruction, I want to examine whether the AI has a way to reach the context it needs.
“Put together a proposal for this project. Make it good.”
From that request, it reviews past proposals, checks the customer’s information and the products they currently use, and builds a proposal informed by similar cases and the company’s pricing structure. If it lacks information needed for a decision, it asks.
If we can reach that point, working with AI will change considerably.
From writing prompts to building the environment
When I first started using generative AI, much of my attention went toward writing good prompts.
How we give instructions still matters, of course. But in my own work, the balance of time I spend on preparation has gradually shifted.
Keep the conditions AI needs to do its work in place, as part of everyday practice.
Take Google Drive.
File names are vague. Customer documents and internal materials are mixed together. Nobody can tell which version is current, or trace the connections between the proposal, the estimate, and the contract.
Simply connecting AI to this does not necessarily give it what it needs. Even if it can search the documents, deciding which ones should inform the current task remains a separate problem.
A person might have a memory to go on: “I think Tanaka put that document together last year.” AI can also be given ways to consult past records, but that history needs to have been recorded and made accessible.
Without it, the conversation starts to look like this:
“No, not that document.”
“The latest version, not the 2024 one.”
“We don’t sell that product anymore.”
“We mustn’t make that proposal to this customer.”
Correcting each point may move the immediate task forward. But unless those corrections are preserved in a form that helps with the next task, the same explanations will be needed again.
Each time that happens, there is something to examine before questioning the model’s performance.
Making the company’s judgment available for reference
Suppose models improve further, becoming better at coding, writing, and research.
General ability alone still cannot fill in what we promised our customers, which prices have been approved, or why we decided against a proposal in the past.
What will make the difference is whether a company enables AI to access the information it needs and use it to make decisions. We must put the information and systems for that in place.
Past proposals and failures, customer conversations, product knowledge, operating rules, prices, and decision criteria. What used to be shared implicitly inside the company needs to become something that can be consulted when needed.
But more information is not automatically better.
Which information takes priority? How should old documents be treated? How should information about different customers be kept separate? How much judgment should AI be allowed to exercise, and where should a person step in to check?
All of that belongs in the design.
Trying to do this work also raises questions for the people involved. Which version is current? Who decides on exceptions? Why did we choose that proposal last time? Things colleagues somehow understood between themselves now need to be made explainable.
Organizing information for AI can also become a way for a company to look again at how it makes decisions.
A world where “make it good” moves the work forward
Eventually, I want work to move forward with instructions as simple as:
“Put together a proposal for that company. Make it good.”
“Make a good summary of this month’s numbers.”
“Handle this change to the specifications well.”
That is the kind of setup I want to build.
Prepare the information and systems behind the work well enough that a few words are sufficient. Share an understanding of what can be delegated and what needs to be checked.
It is close to asking a colleague who knows the company, the customer, and the history of the work. They can move ahead without having every step explained, and come back to discuss the points they are unsure about.
I want to move toward being able to delegate work to generative AI in that way, too.
So lately, I have been looking at how “good” the AI’s deliverables actually are when all I say is “make it good.”
Did the result meet expectations? Were the necessary checks made? Did we avoid having to repeat the same explanations?
Where something failed to get across, I revisit both the wording of the instruction and the context behind it.
And now I am also organizing that setup so I can share this ability to “make it good” with the rest of the team. By sharing the information, decision criteria, and tools, I am building an environment in which those few words work for their tasks as well.
When one person tries something and it works, the next person should be able to use it from the start. Someone’s trial and error makes a colleague’s work a little easier. That is how I want to widen the reach of “make it good.”
To make that “good” more reliable, and make it available to more people, I am tinkering with the harness again today.