Context Engineering Beats Better Prompts at Making Enterprise AI Results Repeatable

The views and opinions expressed are those of Johnny Leuthard and do not represent the official policy or position of any organization.

Writing a sharper prompt is the first move when AI output disappoints, and for work that has to run twice, it barely helps. Reliability comes from the information a model sees at each step and the order it arrives in. Set that up once and the same task holds its answer across every run.

Johnny Leuthard is an Information Security Engineer at Wells Fargo. He's spent more than 15 years in security engineering, much of it focused on privileged access management inside large financial institutions. Over the past two years he's built a set of AI-assisted workflows to take on the slow, repetitive parts of his work, an effort that started with solving his own problems and grew from there. The workflows come down to one rule.

"Only give the model the context it needs, when it needs it. Walk it through methodically, one step at a time, and you really shrink your token usage," says Leuthard. A model given everything at once usually returns an answer that needs fixing. Each fix sends the whole thing back, and the model reads it all again, including the parts that didn't change. Usage is priced by the token, so every rewrite raises the bill.

A narrow path

Leuthard assigns the model one small task at a time. "A lot of people think of a constraint as telling the model 'don't do this.' It's also guiding it down a narrow path," he says. One task might be to research a single question and report back.

Each one finishes before the next starts. "It's like baking a cake. You tell it to get the ingredients. That's step one. That's all you do," he notes. Preheating the oven comes next, then mixing. Sending the whole recipe in one instruction produces what he calls context bloat. The model holds so much information that it loses the thread and starts dropping details.

Leuthard learned the approach through interpretable context methodology, a documented method for breaking AI work into small steps. It keeps each step and its context in a set of files and folders the user controls, so the whole setup lives outside any single chat window. "It's still prompts. It's a bunch of mini prompts, but it's how you structure it, how you use it, how you control it," he adds.

Predictable on purpose

Leuthard points to a common frustration with image tools. "You ask for a picture of a house, then say you wanted the door red, and it redoes the whole thing into something completely different," he says. In a controlled workflow, the fixed details are set in advance, down to the style of house and the number of windows. Change the door and only the door changes, while the rest of the image comes back untouched.

The same idea links one stage to the next. Each stage writes its result to a file, and that file becomes the input for the following one. Leuthard borrows a term from software to describe the setup. "Those are microservices. It does its one task and nothing else," he says. This modular design keeps large software maintainable.

Updates take almost no work. "If something changes, all you've got to do is change that one little piece," Leuthard adds. Each stage still hands the next one the same kind of input, so nothing downstream has to be rebuilt. For anyone still working one prompt at a time, the surrounding structure is where the value sits.

Clean data first

Even a carefully built workflow can't rescue a weak data foundation. "AI can only work with clean data," Leuthard adds. One person can keep their own inputs in order. A team is where it gets complicated. Everyone has to agree on where results go and what format they take. Then a shared system can read all of it and feed the dashboards and reports the business runs on.

As models grow more capable, the engineer's job shifts toward oversight. The work becomes designing what the model reads and reviewing what it returns. A model will present a wrong answer with full confidence, and the only safeguard is a person paying attention. "I've had some pretty bad instances where it made stuff up on me. You've got to be careful," says Leuthard.